Simple Error Detection Methods for

Hardware Implementation of
Advanced Encryption Standard
Chih-Hsu Yen and Bing-Fei Wu, Senior Member, IEEE
AbstractIn order to prevent the Advanced Encryption Standard (AES) from suffering from differential fault attacks, the technique of
error detection can be adopted to detect the errors during encryption or decryption and then to provide the information for taking further
action, such as interrupting the AES process or redoing the process. Because errors occur within a function, it is not easy to predict the
output. Therefore, general error control codes are not suited for AES operations. In this work, several error-detection schemes have
been proposed. These schemes are based on the i 1. i cyclic redundancy check (CRC) over G12
8
, where i 2 f4. 8. 16g.
Because of the good algebraic properties of AES, specifically the MixColumns operation, these error detection schemes are suitable
for AES and efficient for the hardware implementation; they may be designed using round-level, operation-level, or algorithm-level
detection. The proposed schemes have high fault coverage. In addition, the schemes proposed are scalable and symmetrical. The
scalability makes these schemes suitable for an AES circuit implemented in 8-bit, 32-bit, or 128-bit architecture. Symmetry also
benefits the implementation of the proposed schemes to achieve that the encryption process and the decryption process can share the
same error detection hardware. These schemes are also suitable for encryption-only or decryption-only cases. Error detection for the
key schedule in AES is also proposed and is based on the derived results in the data procedure of AES.
Index TermsAdvanced encryption standard, error control code, CRC, differential fault attacks.

1 INTRODUCTION
T
HE Advanced Encryption Standard (AES) [10], the
successor to the Data Encryption Standard (DES), was
finalized in October 2000 by the US National Institute of
Standards and Technology (NIST), when the Rijndael
algorithm [12] was adopted. The data block size of AES is
128-bit and the key size can be 128-bit, 192-bit, or 256-bit. In
AES, although the data block is 128-bit, all operations are
byte-oriented over G12 or G12
8
. Therefore, several
kinds of AES implementations have been discussed. In
general, three main types of AES implementations have
been discussed, 8-bit, 32-bit, or 128-bit architecture. Each
architecture has its own applications. Feldhofer et al. [6]
designed an 8-bit AES chip to provide security for radio
frequency identification (RFID). Satoh et al. [13] introduced
a 32-bit implementation of AES. Mangard et al. [9] proposed
a scalable architecture for AES, which could process 128-bit
data or 32-bit data, depending on the number of Sbox.
The hardware implementation of AES would be coun-
tered by some side-channel attacks, such as Differential
Fault Attacks (DFA) or Differential Power Analysis (DPA).
Differential fault attacks was originally proposed by Biham
and Shamir [4]. Theses side-channel attacks actually
threaten the security of several cryptosystems because they
are practical for a crypto module. The idea of DFA is to
apply the differential attacks to a crypto module or a crypto
chip. The cryptanalyst injects errors by using microwave or
ionizing techniques during the encryption or decryption
process. These errors cause the encryption results to differ
from the correct results; hence, the cryptanalyst will receive
the difference of outputs. Therefore, such differential
attacks may be carried out in the real world. Dusart et al.
[5] broke the 128-bit AES under the assumption that you can
physically modify the hardware AES device. This attack
required 34 pairs of differential inputs and outputs to
obtain the final round key. Piret and Quisquater [11] broke
AES with two erroneous ciphertext under the assumption
that the errors occur between the antepenultimate and the
penultimate MixColumns.
To avoid the possibility of suffering such attacks, error
detection can be considered while implementing a cipher.
In 2002, Karri et al. [7] proposed a general error detection
method, called concurrent error detection (CED), for several
symmetric block ciphers including RC6, MARS, Serpent,
Twofish, and Rijndael. CED requires an inverse operation to
check whether errors have occurred in calculations or not
and has three levels: the operation level, the round level,
and the algorithm level. Taking an operation-level CED in
AES as an example, the InvSubBytes is required to detect
the errors occurring in SubBytes and vice versa. This
method has very high fault coverage, but it is time-
consuming and high hardware cost because inverse opera-
tions are required. In 2003, Karri et al. [8] proposed a parity-
based detection technique for general substitution-permu-
tation block ciphers. However, the size of the table, required
by the substitution box, is enlarged. In addition, the paper
did not address the error detection techniques for some
specific functions, such as MixColumns in AES. In 2004,
Wu et al. [14] applied the structure of [8] to AES and used
720 IEEE TRANSACTIONS ON COMPUTERS, VOL. 55, NO. 6, JUNE 2006
. The authors are with the Department of Electrical and Control Engineerng,
National Chiao Tung University, 1001 Ta Hsueh Rd., Hsinchu, Taiwan
300, ROC. E-mail: {zsyian, bwu}@cssp.cn.nctu.edu.tw.
Manuscript received 19 Jan. 2005; revised 27 Aug. 2005; accepted 7 Sept.
2005; published online 21 Apr. 2006.
For information on obtaining reprints of this article, please send e-mail to:
tc@computer.org, and reference IEEECS Log Number TC-0014-0105.
0018-9340/06/$20.00 2006 IEEE Published by the IEEE Computer Society
one-bit parity for a 128-bit data block. The method of Wu
et al. [14] can let the parity pass through the MixCol-
umns. Bertoni et al. [1] used an error detection code of 16-
bit parity for a 128-bit data block. To be precise, this
approach uses one-bit parity for each byte and, thus, can
detect all single errors and perhaps all odd errors. In [2],
Bertoni et al. used the error detection scheme in [1] not
only to detect errors but also to locate errors. In 2004,
Bertoni et al. [3] implemented the model proposed in [2].
The introduction of the mode into AES brought the
performance 18 percent overhead of area and 26 percent
decreasing of throughput. According to the results given
in [1], their approach was able to detect most cases of
multiple faults. However, this approach is asymmetrical,
between MixColumns and InvMixColumns, because the
parity prediction of InvMixColumns is more complex
than that of MixColumns. Therefore, two circuits are
required to predict the parity while merging the encryption
and the decryption. Besides, the detection technique for
SubBytes doubled the table size of SubBytes in AES,
from 256 to 512 bytes. In addition, it cannot be easily
applied to an AES implementation of 8-bit architecture
because the parity prediction of MixColumns (InvMix-
Columns) requires information from other bytes and other
parities.
This work proposes several error-detection schemes for
AES. They are based on the i 1. i cyclic redundancy
check (CRC) over G12
8
, where i 2 f4. 8. 16g is the
number of bytes contained in the message. The proposed
schemes easily predict the parity of an operations output.
Because AES is byte-oriented and its constants are
ingeniously designed, the parity of the output can be
predicted from a linear combination of the parity of the
input. In most cases, the parity is the summation of the
input data; also, the proposed schemes are highly scalable
and are suitable for 8-bit, 32-bit, or 128-bit architecture. This
is important because many AES designs are in an AES
hardware designed as either 8-bit or 32-bit architecture.
Another advantage of the proposed approaches is that the
parity calculation between the encryption and the decryp-
tion is symmetric because the parity generation in encryp-
tion is quite similar to the one in decryption. This will bring
some benefits while integrating encryption and decryption
into one circuit.
This paper is organized as follows: In Section 2, the AES
algorithm is briefly described and the notations used
throughout are defined. In Section 3, our proposed error
detection schemes for AES are described. Derivation of
error detection for each operation, including SubBytes,
ShiftRows, MixColumns, and AddRoundKey, is ex-
plained, as well as the design of the key schedule. The
undetectable errors of each proposed method are theoreti-
cally analyzed in Section 4, while, in Section 5, the
realization issues of three levels, operation level, round
level, and algorithm level, are described. In Section 6,
advantages and comparisons between this work and other
research studies are discussed and, in Section 7, the
detection capability of each scheme is simulated. Finally,
our conclusions are offered in Section 8.
2 AES ALGORITHM
The AES [10] consists of two parts, the data procedure and
the key schedule. The data procedure is the main body of the
encryption (decryption) and consists of four operations,
(Inv)SubBytes, (Inv)ShiftRows, (Inv)MixColumns,
and (Inv)AddRoundKey. During encryption, these four
operations are executed in a specific orderAddRoundKey,
a number of rounds, and then the final round. The number
of rounds is 10, 12, or 14, respectively, for a key size of
128 bits, 192 bits, or 256 bits. Each round is comprised of the
four operations and the final round has SubBytes,
ShiftRows, and AddRoundKey. The decryption flow is
simply the reverse of the encryption, and each operation is
the inverse of the corresponding one in encryption. In the
data procedure, the 16-byte (128-bit) data block is rear-
ranged as a 4 4 matrix, called state o,
o
:
0
:
4
:
8
:
12
:
1
:
5
:
9
:
13
:
2
:
6
:
10
:
14
:
3
:
7
:
11
:
15
2
6
6
4
3
7
7
5
. 1
where :
i
denotes the ith byte of the data block. In this
context, o denotes the input of an operation and T denotes
the output. AES is operated in two fields, G12 and
G12
8
. In G12, addition is denoted by , and multi-
plication is denoted by . Similarly, the two symbols, and
, denote addition and multiplication in G12
8
.
2.1 SubBytes
Two calculations, the G12
8
inversion and the affine
transformation, are involved in this operation. SubBytes
substitutes each byte :
i
of the data block by
t
i
:
1
i
63. 2
where :
1
i
is the inverse of the input byte, :
i
2 G12
8
, is
an 8 8 circulant matrix of a constant row vector
1 0 0 0 1 1 1 1 over G12, and 63 (the Courier font
number representing a hexadecimal value in this paper)
belongs to G12
8
. :
1
i
is a matrix-vector multiplication
over G12.
2.2 ShiftRows
The ShiftRows operation only changes the byte position
in the state. It rotates each row with different offsets to
obtain a new state as follows:
:
0
:
4
:
8
:
12
:
1
:
5
:
9
:
13
:
2
:
6
:
10
:
14
:
3
:
7
:
11
:
15
2
6
6
4
3
7
7
5
ShiftRows
!
:
0
:
4
:
8
:
12
:
5
:
9
:
13
:
1
:
10
:
14
:
2
:
6
:
15
:
3
:
7
:
11
2
6
6
4
3
7
7
5
. 3
The first row is unchanged, the second row is left circular
shifted by one, the third row is by two, and the last row is
by three.
2.3 MixColumns
The MixColumns operation mixes every consecutive four
bytes of the state to obtain four new bytes as follows:
YEN AND WU: SIMPLE ERROR DETECTION METHODS FOR HARDWARE IMPLEMENTATION OF ADVANCED ENCRYPTION STANDARD 721
:
0
:
4
:
8
:
12
:
1
:
5
:
9
:
13
:
2
:
6
:
10
:
14
:
3
:
7
:
11
:
15
2
6
6
4
3
7
7
5
MixColumns
!
t
0
t
4
t
8
t
12
t
1
t
5
t
9
t
13
t
2
t
6
t
10
t
14
t
3
t
7
t
11
t
15
2
6
6
4
3
7
7
5
. 4
Let :
i
, :
i1
, :
i2
, and :
i3
represent every consecutive four
bytes, where i 2 f0. 4. 8. 12g. Then, the four bytes are
transformed by
t
i
t
i1
t
i2
t
i3
2
6
6
4
3
7
7
5

02 03 01 01
01 02 03 01
01 01 02 03
03 01 01 02
2
6
6
4
3
7
7
5
:
i
:
i1
:
i2
:
i3
2
6
6
4
3
7
7
5
. 5
Each entry of the constant matrix in (5) belongs to G12
8
,
hence (5) is a matrix-vector multiplication over G12
8
.
2.4 AddRoundKey and Key Expansion
Each round has a 128-bit round key which is segmented
into 16 bytes /
i
as (1); the AddRoundKey operation is
simply an addition,
t
i
:
i
/
i
. where 0 i 15. 6
The key expansion expands a unique private key as a key
stream of 4i 4 32-bit words, where i is 10, 12, or 14. The
private key is segmented into Nk words according to the key
length, where NK is 4, 6, or 8 for a 128-bit, 192-bit, or 256-bit
cipher key, respectively. As Fig. 1shows, then, it generates the
ithword(32 bits) by EXORing the i Nkth wordwitheither
the i 1th word or the conditionally transformed i 1th
word, where NK i 4i 3. The i 1th word is con-
ditionally transformed by RotWord, SubBytes and EXOR-
i ng wi t h Rconi,Nk f02
bi,Nkc
. 00. 00. 00g, where t he
polynomial presentation of 02
bi,Nkc
is r
bi,Nkc
over G12
8
.
Finally, the key stream is segmented into several round keys
which are involved in the AddRoundKey operation.
3 ERROR DETECTION TECHNIQUES
The parts in decryption can be yielded in a similar way;
hence, the following context only addresses the error
detection in encryption. The differential faults attacks need
differential inputs and outputs to attack a cryptosystem;
hence, it is assumed that the states and round keys are
polluted by additive errors, as shown in Fig. 2. In this work,
one operation is the smallest granule for designing error
detection. In Fig. 2, the errors are assumed to be induced
between the previous operation and the current operation.
If the errors occur in the output of the previous operation,
the erroneous input of the current operation will be treated
as a different state. Actually, this situation only exists in the
first round or in the first operation. The assumed error
model is logical, even in the case where the errors occur
during the operation. Because each operation of AES is
invertible, one unique error block c would exist for an
erroneous output T such that T )o c, where ) denotes
any operation in AES.
This paper adopts a systematic i 1. i cyclic redun-
dancy check (CRC) over G12
8
to detect errors occurring
duringencryption, where i 2 f4. 8. 16gis thenumber of bytes
contained in the message. The generator polynomial is
qr 1 r. 7
where the coefficients of (7) are over G12
8
. Giving a
message :r of degree i 1, a systematic codeword,
generated by qr, can be obtained from the following two
steps:
1. Obtain the remainder jr from dividing r:r by
the generator polynomial qr. The remainder jr is
a scalar j here because the degree of qr is one.
2. Combine jr and r:r to obtain the codeword
polynomial,
jr r:r j :
0
r :
1
r
2
:
i1
r
i
.
where j. :
i
2 G12
8
.
8
In Step 1, while qr is 1 r, the remaining jr is the
summation of all coefficients of the message,
jr
X
i
i0
:
i
. 9
722 IEEE TRANSACTIONS ON COMPUTERS, VOL. 55, NO. 6, JUNE 2006
Fig. 1. The block diagram of key expansion in AES.
Fig. 2. The error model assumed in this work. The solid line part appears
in every operation and the dotted line part appears in some operations.
Therefore, the parity of a message may be obtained by
calculating the summation of the input message over G12
8
.
Assume that the received polynomial tr is
tr t
0
t
1
r t
2
r
2
t
i
r
i
. t
i
2 G12
8
. 10
The detection scheme checks whether the syndrome equals
zero or not, where syndrome n is
n
X
i
i0
t
i
. 11
If the syndrome equals zero, then it is assumed that no
errors have occurred; otherwise, errors did occur.
In the channel coding field, it is assumed that the
message :r is transmitted over a noisy channel. The
channel does not modify the message if no errors occur.
Therefore, it is easy to predict that t
0
is identical to j, with t
0
being used to detect the errors. However, as shown in Fig. 3,
the message, o f:
0
. :
1
. . . . . :
i1
g, is transformed into
another message, ft
1
. t
2
. . . . . t
i
g, by an AES operation;
hence, t
0
cannot be obtained instinctively. Therefore, this
paper investigates the function, predicting t
0
from j as
shown in Fig. 3, for each operation to make error detection
possible in AES.
This work applies an i 1. i CRC to AES, where
i 2 f4. 8. 16g. In the case where, i 16, a 128-bit AES state
is treated as a message; hence, only one parity is generated
for a 128-bit data block. When i 4, the error detection is
designed to check each column of the output state. In other
words, four 4-byte column vectors in an AES state,
ft
4,1
. t
4,2
. t
4,3
. t
4,4
g, 0 , 3, are checked separately.
Therefore, four parities are required for a 128-bit data block
when i 4. For i 8, two parities are required for a
128-bit data block. The following context addresses the two
cases, i 16 and i 4, because the 9. 8 CRC for the AES
algorithm can be constructed under similar conditions to
the 17. 16 or 5. 4 CRC for AES.
3.1 In SubBytes
In this paper, two implementation types of SubBytes are
considered. The first type uses one table instead of the
G12
8
inversion and the affine transformation. The second
type separately calculates the G12
8
inversion and the
affine transformation and the implementation of the
G12
8
inversion is not limited to the look-up-table method
or the combinational logical circuit. In this paper, the first
type is named united SubBytes and the second type is
separated SubBytes.
For united SubBytes, it is assumed that both the Sub
Bytes circuit and the InvSubBytes circuit are imple-
mented in a chip. Error detection is achieved by feeding
the output of SubBytes into InvSubBytes, then
comparing the input of SubBytes and the output of
InvSubBytes, and vice versa, as Fig. 4 shows. If both
are identical, then it is concluded that no errors have
occurred. Otherwise, the errors did occur. This error
detection method may be time-consuming, if only the
SubBytes operation is considered. However, in practical
terms, normal encryption could be further processed,
without waiting for the error detection result, because
SubBytes is either the first operation or the second
operation in each round. In other words, the operation
after SubBytes, such as ShiftRows, MixColumns, or
AddRoundKey, may continue, when the output of the
round would be intercepted if errors are detected in
SubBytes.
If separated SubBytes is adopted, error detection must be
applied separately to the G12
8
inversion and the affine
transformation. Considering the error detection for the
G12
8
inversion first, there are two schemes proposed
herein. Similarly to Fig. 4, the first scheme detects errors by
using the relationship of the mutual inverse. However, the
computation of the G12
8
inversion is identical for both
SubBytes and InvSubBytes; hence, this scheme does not
require the encryption and decryption circuits to simulta-
neously exist in one chip. It can be used with the
encryption-only or decryption-only hardware.
The second scheme is the i 1. i CRC and assumes
that the G12
8
inversion is implemented in look-up-table
approach. Instead of the inverse value of a giving input, the
exclusive value of the giving input and its inverse is stored
in the table. Therefore, giving an input c 2 G12
8
, the
value, u c c
1
, is obtained from the table and then the
input c is added to u to yield c
1
, as the marked block in
Fig. 5. The error is detected by the syndrome obtained by
the dashed line in Fig. 5. In this diagram, no errors are
introduced, hence the syndrome is zero.
For one G12
8
inversion, according to Fig. 3 and the
error model given in Fig. 2, the errors induce a fault at the
input of the G12
8
inversion, as shown in Fig. 6. Suppose
that the byte :
i
is changed into another byte :
0
i
by adding the
error c
0
. Then, the syndrome used to detect errors is
calculated as
:
i
c
1
t
i1
t
i1
t
1
i1
c
0
c
1
. 12
The one-byte structure of Fig. 5 could be extended to the
4-byte, 8-byte, or 16-byte structure. Taking the 16-byte
YEN AND WU: SIMPLE ERROR DETECTION METHODS FOR HARDWARE IMPLEMENTATION OF ADVANCED ENCRYPTION STANDARD 723
Fig. 3. The block diagram of the error detection in this paper.
Fig. 4. The error detection for united SubBytes.
structure into consideration, the input state is denoted as
o f:
0
. :
1
. . . . . :
15
g and then the parity j is
P
15
i0
:
i
from (9).
According to (12) and Fig. 3, the parity of the output
parity t
0
could be predicted by
X
15
i0
:
i

X
15
i0
t
i1
t
1
i1
. 13
and the syndrome is
t
0

X
15
i0
t
i1
.
)
X
15
i0
t
i1
j
X
15
i0
t
i1
t
1
i1
.
14
If no errors have occurred, the value t
1
i1
will equal :
i
.
Therefore, the syndrome (14) is zero.
In this paper, all ShiftRows, MixColumns, and
AddRoundKey are protected by error detection code.
However, the detection technique of SubBytes is varied
with its implementation. According to the error detection
scheme for SubBytes, three proposed architectures for
AES are denoted by united-SubBytes detection (USBD, hybrid-
SubBytes detection (HSBD), and parity-based-SubBytes detec-
tion(PbSBD), as shown in Fig. 7.
For the affine transformation, error detection is achieved
by the i 1. i CRC, where i 2 f4. 8. 16g. Considering i
16 first, and according to (9), the parity j of an input state,
o f:
0
. :
1
. . . . . :
15
g, where :
i
2 G12
8
, is generated by
j
X
15
i0
:
i
. 15
The output state is denoted as T ft
0
. t
1
. . . . . t
16
g. From
(2) and Fig. 3, t
i1
is :
i
63, where 0 i 15. The
hexadecimal constant 63 will be eliminated after taking
summation of the output state Tnt
0
, i.e.,
X
i1
i0
t
i1

X
i1
i0
:
i
63
X
15
i0
:
i
j. 16
Therefore, t
0
can be predicted by (16) with input parity j. If
no errors occur, the syndrome n must be zero,
n
X
16
i0
t
i
0. 17
In the case of 5. 4 CRC or 9. 8 CRC, (16) also holds.
3.2 In ShiftRows
From (3), the ShiftRows operation simply rotates the
input state o, but does not alter the value of :
i
. Therefore, t
0
may be directly predicted by
P
i
i0
:
i
in the case of i 16.
Similarly, the ShiftRows operation is error free if the
syndrome is zero
X
16
i0
t
i
0. 18
When i 4, because each column of the output state would
be detected, the four parities j
,
, where 0 , 3, are
j
0
:
0
:
5
:
10
:
15
.
j
1
:
4
:
9
:
14
:
3
.
j
2
:
8
:
13
:
2
:
7
.
j
3
:
12
:
1
:
6
:
11
;
hence, the t
,.0
for eachoutput message ft
4,1
. t
4,2
. t
4,3
. t
4,4
g
is j
,
. The case of i 8 is analogous to the case of i 4.
724 IEEE TRANSACTIONS ON COMPUTERS, VOL. 55, NO. 6, JUNE 2006
Fig. 5. The block diagram of one G12
8
inversion with the error
detection.
Fig. 6. An error is injected into the input state after entering the G12
8

inversion.
Fig. 7. The three proposed architectures for AES.
3.3 In MixColumns
The behavior of the MixColumns operation is more
complex because each byte in the input state o influences
four bytes in the output state T. However, because of the
ingenious design of the matrix coefficients, it is also possible
to apply the i 1. i CRC directly, where i 2 f4. 8. 16g.
The MixColumns operation works as follows:
t
4,1
t
4,2
t
4,3
t
4,4
2
6
6
4
3
7
7
5
|{z}
T
0

02 03 01 01
01 02 03 01
01 01 02 03
03 01 01 02
2
6
6
4
3
7
7
5
:
4,
:
4,1
:
4,2
:
4,3
2
6
6
4
3
7
7
5
|{z}
o
0
. where 0 , 3.
19
From (19), it is yielded that the summation of vector T
0
equals that of vector o
0
.
X
3
/0
t
4,/1
02 01 01 03:
4,

03 02 01 01:
4,1

01 03 02 01:
4,2

01 01 03 02:
4,3
.
:
4,
:
4,1
:
4,2
:
4,3
.

X
3
/0
:
4,/
.
20
Therefore, when the 5. 4 CRC is applied, the output parity
t
,.0
of the ,th column vector may be directly predicted from
the ,th column vector of the input state by
P
3
/0
:
4,/
.
Similarly, in the case i 16, t
0
is predicted by
t
0

X
3
,0
X
3
/0
t
4,/1
.

X
3
,0
X
3
/0
:
4,/
.

X
15
i0
:
i
.
Because the summation of 02, 01, 01, and 03 is 01, (20)
can be satisfied for the 17. 16, 9. 8, or 5. 4 CRC. The
coefficients of InvMixColumns display an identical phe-
nomenon. The summation of the four coefficients used in
decryption, 0B, 0D, 09, 0E, is also 01. Therefore, t
0
or t
,.0
can be predicted in the same way as that of MixColumns.
3.4 In AddRoundKey
Discussing the case i 16 first, it is assumed that each
round key already has a parity; hence, the round key is
represented as f/
0
. /
1
. . . . . /
16
g, where /
0

P
15
i0
/
i1
is the
parity and f/
1
. . . . . /
16
g is the normal round key. The
AddRoundKey operation only adds the input state with a
normal key 1 f/
1
. /
2
. . . . . /
16
g to yield the output state as
follows:
T o 1. 21
We apply the summation operation to (21) to obtain
X
15
i0
t
i1

X
15
i0
:
i

X
15
i0
/
i1
j /
0
. 22
Accordingly, t
0
may be obtained from j /
0
. The parities
for i 4 or i 8, j
,
, are calculated in the same way;
however, the round key must also have four or two parities.
3.5 In the Key Expansion
The i 1. i CRC is also adopted in key expansion, where
i 2 f4. 8. 16g. However, the 5. 4 CRC is always used in the
interior of the key expansion. The key expansion and the
error detection scheme are jointly depicted in Fig. 8, where
the decision blocks are removed from Fig. 1 for a simple
description of error detection, as the conditions only
determine where the error detection is applied, not how it
is designed.
In this key expansion, with error detection, one word
contains five bytes and the symbol of a word is denoted by
W
0
i Wi k parity, where k is a catenation symbol. At
first, the parities of the first Nk words, where Nk 2 f4. 6. 8g,
are obtained by the generator 1 r, i.e., the parity j
i
of
Wi n
i.0
n
i.1
n
i.2
n
i.3
is
j
i
n
i.0
n
i.1
n
i.2
n
i.3
. 23
Then, the Nk-pair parities and messages formnewNk words,
W
0
0. W
0
1. . . . , and W
0
Nk 1. The new words are succes-
sively put into the Nk shift blocks, from W
0
i Nk to
W
0
i 1, at the top of Fig. 8, after which, the key expansion
starts. A 128-bit round key and its one-byte parities are
collected after each period of four shifts. If 17. 16 CRC is
chosen for AES, the one-byte parity of a round key is
obtained by summing the four parities of output words. If
5. 4 CRC is chosen, then the four parities are kept.
In the key expansion, the RotWord rotates the byte order
of Wi 1; hence, the parity is the same as that of W
0
i 1.
For the SubWord operation because it is a function which
executes SubBytes on each byte of input, the error
detection scheme is the same as that in SubBytes,
YEN AND WU: SIMPLE ERROR DETECTION METHODS FOR HARDWARE IMPLEMENTATION OF ADVANCED ENCRYPTION STANDARD 725
Fig. 8. The error detection scheme for key expansion.
described in Section 3.1. However, in the case of united
SubBytes being used, the parity must be calculated
separately.
For the EXOR operation with Rconi,Nk, the error
detection is achieved by EXORing the parity of temp and
that of Rconi,Nk, where Rconi,Nk f02
bi,Nkc
. 00. 00. 00g.
The parity of Rconi,Nk equals 02
bi,Nkc
due to the three
bytes of zero value in Rconi,Nk. At the end of the key
expansion, the parity t
0
is the EXOR of the parity of current
data and the parity of W
0
Nk 1.
3.6 More Details for 5. 4 CRC
Although the 5. 4 CRC has four parities, it is possible for
only one parity to be used in realization of this scheme. AES
can be implemented in a 32-bit structure, i.e., one column of
a state is processed once in every round. In this structure,
the position of ShiftRows must be shifted above the
SubBytes operation. After ShiftRows, each column
passes through the identical calculations, SubBytes,
MixColumns, and AddRoundKey; the parity generation,
or the syndrome calculation for each column, are also
identical, so only one circuit is required.
4 UNDETECTABLE ERRORS
Even though the AES algorithmpropagates the errors during
encryption, the error coverage can be also analyzed mathe-
matically. Actually, only the MixColumns and SubBytes
operations cause numerous erroneous bits when a single-bit
error is injected, when ShiftRows or AddRoundKey do
not change the bit number of the errors. Several assump-
tions are made, as follows:
1. The error model is considered as Fig. 2.
2. All nonzero error block over G12
8i1
have the
same probability, where i 2 f4. 8. 16g.
3. Each operation has the same error injection
probability.
4.1 The Undetectable Errors in SubBytes
Because SubBytes is invertible, all errors injected into
input can be detected by InvSubBytes and vice versa.
Therefore, the united SubBytes, has 100 percent fault cover-
age. In separated SubBytes, both operations, the G12
8

inversion and the affine transformation, have their own

16
1 and the
percentage of undetectable errors is
2
8

16
1
2
8

17
0.4%. When
the 5. 4 CRC is applied to a 128-bit data block, the total
number of undetectable nonzero errors is 2
8

4
1
4
and
the percentage is
2
8

4
1
2
8

5

4
100% 2.56 10
8
%. Simi-
larly, the percentage of undetectable errors for the
9. 8 CRC is 0.16 10
2
%.
The affine transformation is detected by i 1. i CRC.
Although five erroneous bits were caused, while injecting a
single-bit error, the error coverage can still be analyzed.
Theorem 1. Given an input state o fj. :
0
. :
1
. . . . . :
i1
g,
where parity j is
P
i1
i0
:
i
, and i 2 f4. 8. 16g, the output state
is T ft
0
. t
1
. . . . . t
i
g, where t
0
is j from (16), and t
i1
,
0 i i 1, is obtained from (2). Introducing an error 1
fc
0
. c
1
. . . . . c
i
g into the state o fj. :
0
. :
1
. . . . . :
i1
g, the
summation of the output T
0
will equal to zero if and only if
P
i
i0
c
i
0.
Proof. Because i is even, the value 63 will be cancelled.
Therefore, the summation of the erroneous output T
0
is
X
i
i0
t
0
i
j c
0

X
i1
i0
:
i
c
i1
.
j
X
i1
i0
:
i
|{z}
0

X
i
i0
c
i
.

X
i
i0
c
i
.
Therefore,
P
i
i0
t
0
i
equals to zero if and only if

P
i
i0
c
i
0 is held. Because the matrix is nonsin-
gular over G12,
P
i
i0
c
i
is zero if and only if
P
i
i0
c
i
is zero. tu
In the i 1. i CRC, the nonzero errors are undetected,
when the equation
P
i
i0
c
i
0 is held, i.e., errors are also
the codewords. According to Theorem 1, all undetectable
errors are also undetected after the affine transformation.
Therefore, while applying the i 1. i CRC to a 128-bit
data block, the percentages of the undetectable errors are
0.4 percent, 0.16 10
2
%, and 2.56 10
8
%, respectively,
for i 16, i 8, and i 4.
4.2 The Undetectable Errors in MixColumns
MixColumns also has a diffusion property. It causes five or
11 erroneous bits while injecting a single-bit error in one
column vector of the input state. However, the coefficients
eliminate the diffusion of errors after summing the erroneous
columnvector of theoutput state. The MixColumns is shown
again below, and it is supposed that each byte of the input
vector is polluted by an error.
t
i1
t
i2
t
i3
t
i4
2
6
6
4
3
7
7
5

02 03 01 01
01 02 03 01
01 01 02 03
03 01 01 02
2
6
6
4
3
7
7
5
:
i
c
i
:
i1
c
i1
:
i2
c
i2
:
i3
c
i3
2
6
6
4
3
7
7
5
. 24
Then, the summation of the column vector t
i1
is
X
3
/0
t
i/1
02 01 01 03:
i
c
i

03 02 01 01:
i1
c
i1

01 03 02 01:
i2
c
i2

01 01 03 02:
i3
c
i3
.

X
3
/0
:
i/
c
i/
.
25
The equation also holds for two or four columns vectors.
726 IEEE TRANSACTIONS ON COMPUTERS, VOL. 55, NO. 6, JUNE 2006
Theorem 2. Giving an input state o fj. :
0
. :
1
. . . . . :
i1
g,
where j
P
i1
i0
:
i
is the checksum of the input state and
i 2 f4. 8. 16g. After MixColumns and the parity prediction
(20), the output state is T ft
0
. t
1
. . . . . t
i
g, where t
0
j, and
the rest is the output of MixColumns. Introducing an error
1 fc
0
. c
1
. . . . . c
i
g into the state o fj. :
0
. :
1
. . . . . :
i1
g,
then the errors of the i 1. i CRC in MixColumns are
undetectable if and only if the summation
P
i
i0
c
i
is zero.
Proof. The syndrome
P
i
i0
t
i
is used to check whether
errors occurred or not. It is assumed that no errors
occurred, if and only if the syndrome is zero. The
summation of the erroneous output state is
X
i
i0
t
0
i
t
0
c
0

X
i
i1
t
0
i
.
From (25), because i is the multiple of four, the above
equation is represented as
X
i
i0
t
0
i
t
0
c
0

X
i
i1
:
i1
c
i
.
t
0

X
i1
i0
:
i
|{z}
0

X
i
i0
c
i
.

X
i
i0
c
i
.
Therefore, the error is undetectable if and only if
P
i
i0
c
i
is zero. tu
From Theorem 2, there are 2
8

16
1 nonzero errors
that are undetectable, when the 17. 16 CRC is applied to a
128-bit data block. This result is the same as those in the
affine transformation described above. Similarly, the total
number of the undetectable errors for the 9. 8 or 5. 4 CRC
is 2
8

4
1
4
or 2
8

8
1
2
, respectively.
4.3 The Undetectable Errors in ShiftRows or
AddRoundKey
ShiftRows does not change the value of the input state, and
AddRoundKey only EXORs the input state with a round key.
Therefore, the undetectable errors are the same as those
analyzed in the affine transformation or MixColumns.
5 DETECTION LEVELS
The proposedscheme maybe usedinoperation-level, round-
level, or algorithm-level error detection. In operation-level
detection, the syndrome is checked at the end of each
operation. Similarly, if the syndrome is obtained at the end
of each round, it is round-level detection. The implementa-
tion of operation-level error detection is easy to figure out.
The syndrome is calculated at the end of each operation
according to the equations derivedinSection 3. However, the
implementation of a round-level detection needs more
ingenuity, when the SubBytes is protected by united
SubBytes. The parity is generated at the end of the SubBytes
or the beginning of the ShiftRows. Then, the parity directly
passes through ShiftRows, and MixColumns because its
value will not be changed after the two operations. Finally,
the parity is EXORed with the key parity. The total path is
shown in Fig. 9. Obviously, the syndrome could then be
checked at the end of the round. In hybrid SubBytes, the
structure for round-level error detection is similar to Fig. 9,
but the parity is generated after the G12
8
inversion.
Because the parity of the state, in the ith round, cannot pass
through the inversion of G12
8
in i 1 round, the parity
must be regenerated in each round. Therefore, united-
SubBytes detection or hybrid-SubBytes detection cannot be
implemented as algorithm-level detection.
However, each operation of parity-based SubBytes is
protected by i 1. i CRC, hence the parity could pass
through a round. Therefore, parity-based SubBytes could be
applied as an operation-level, round-level, or algorithm-
level error detection.
6 FEATURES AND COSTS
6.1 Scalability
In Section 3, it was found that the three error detections,
i 1. i CRC, where i 2 f4. 8. 16g, had similar structures.
The calculations of parities or syndromes were all based on
Byte-EXOR (B-EXOR) operation and the length of the
message was a multiple of four bytes. Therefore, the
proposed approach is scalable with practical hardware
design; in other words, the three CRCs can be applied to an
AES implementation of an 8-bit, 32-bit, or 128-bit structure.
In general, the portable devices are more probable to
encounter DFA than a nonportable device. Therefore, the
scalability of error scheme is good for practical purposes
because 8-bit and 32-bit architectures are most commonly
used in portable applications, such as cell phones, Smart-
Card, or RFID tag.
The approach proposed by Bertoni et al. [1] cannot be
easily scaled down into the 8-bit architecture because the
parity of :
i
requires the information from :
i1
and :
i2
.
However, this work can easily be applied to an 8-bit, 32-bit,
or 128-bit AES architecture. The syndrome generation is
similar to parity generation. Fig. 10 shows a block diagram
of (17) and (16) for 8-bit AES architecture. While 16 bytes t
i
are obtained, the syndrome n is obtained immediately,
where the initial value of parity registers as a zero byte. The
ShiftRows, MixColumns, or AddRoundKey have similar
structures to Fig. 10, but the matrix transformation, , is not
required. The 32-bit or 128-bit AES can also be implemen-
ted, based on the concept in Fig. 10.
The 32-bit architecture is the most flexible structure from
the point of error detection because it could use 17. 16,
YEN AND WU: SIMPLE ERROR DETECTION METHODS FOR HARDWARE IMPLEMENTATION OF ADVANCED ENCRYPTION STANDARD 727
Fig. 9. The proposed scheme under round-level error detection.
9. 8, or 5. 4 CRC to achieve the error detection objective.
No matter which one is selected, it is possible that only a
one-byte register is required to store the parities. However,
the input must be a one-column vector, defined in AES;
thus, (20) may be used to detect faults for a one-column
calculation.
6.2 Symmetry
From Fig. 10, it can be seen that the proposed scheme is
symmetric in both encryption and decryption. This has the
advantage of the encryption and decryption being inte-
grated into one chip. However, the scheme proposed by
Bertoni et al. [1] is asymmetrical in MixColumns and
InvMixColumns. As shown in Table 1, the output parity
prediction of InvMixColumns is more complex than that
of MixColumns.
6.3 Costs
While introducing proposed error detection schemes into
AES, the hardware cost required by those schemes is
evaluated through their computational complexity. Error
detection consists of two partsthe parity and syndrome
generation. Discussing the cost in parity generation first, in
our proposed schemes, the parity requires only the EXOR
operation. A total of i 1
16
i
Byte-XORs (B-EXOR) is
required to calculate the parity of the input for the proposed
approach. Taking the 5. 4 CRC for a 128-bit data block as
an example, one checksum of an input message is generated
by three B-EXORs and a total of 12 B-EXORs for four
parities. However, united SubBytes uses InvSubBytes to
check error, so no parity generation is required. In hybrid
SubBytes, the i 1. i CRC is applied to the affine
transformation; 15, 14, or 12 B-XORs are required to
produce the parities for i of 16, 8, or 4, respectively. In
the method proposed by Bertoni et al. [1], 16 7 bit-EXORs
(b-EXOR) were required to obtain 16 one-bit parities for an
AES state. In [7], they used the inversion operation to detect
the errors; hence, no parities were paid for. However, the
hardware of parity generation is minor because the parity
generation is required to perform at the beginning of the
parity-based detection is applied. In PbSBD, because the
parity can pass through each operation along with predict-
ing the parity, the parity generation only performs once. In
USBD and HSBD, the parity must be regenerated in
SubBytes of each round; nevertheless, only one circuit of
parity generation is required when one round is imple-
mented to achieve AES computing. In the approach of
Bertoni et al. [1], the parity can also pass through the round;
hence, one circuit of parity generation is required.
As regards the cost of the syndrome generation and
parity prediction, it varies from operation to operation.
United SubBytes uses the InvSubBytes to detect errors. In
hybrid SubBytes, the G12
8
inversion is used to self-check
errors; the i 1. i CRC is used to detect errors of affine
transformation. According to (17), 16 B-EXORs are required
to obtain the syndrome for every i 1. i CRCs. However,
the execution number of affine multiplication to predict
parity, (16), depends on i; the number is one, two, or four
when i is 16, 8, or 4, respectively. For parity-based SubBytes,
the cost in affine transformation is the same as that in hybrid
SubBytes. However, the G12
8
inversion also uses i
1. i CRC; according to (14), 32 B-EXORs are required (note
that the t
i1
t
1
i1
in (14) is obtained from a table, not
requiring EXOR calculation). In ShiftRows and MixCol-
umns, no prediction functions are necessary and the
syndrome is obtained by summing all output byte and the
parity. Therefore, in the two operations, 16 B-EXORs are
required. In AddRoundKey, the one, two, or four one-byte
parities of a round key are involved in the parity prediction,
requiring extra B-EXORs to be paid for. The results
summarized in Table 1 are the cost of the operation-level
detection, i.e., the error detection is at the end of every
operation. If round-level or algorithm-level are chose, only
728 IEEE TRANSACTIONS ON COMPUTERS, VOL. 55, NO. 6, JUNE 2006
Fig. 10. The block diagram of error detection for 8-bit AES architecture.
TABLE 1
The Cost of Syndrome Generation and Parity Prediction in Each AES Operation in the Operation-Level Detection
B-EXOR = 8 b-EXORs, b-EXOR = bit EXOR operation, EN = encryption, DE = decryption, and AM = affine multiplication.
the cost of parity prediction is required in every operation
and the cost of syndrome generation is only paid at the end
of each round or of the AES algorithm, respectively.
The costs of Bertoni et al.s [1] approach are also varied in
each operation. The SubBytes requires extra i 256-byte
memory spaces to predict the parity, where i is dependent
on the implementation of the AES. Taking an AES
implemented in a 32-bit structure as an example, four bytes
are calculated in parallel, thus four tables are required. The
size of a table with error detection, in [1], is a double of that
in AES, so a total of 512 bytes is for one table, i.e., 256 extra
bytes are caused for one table. The 256 extra bytes are
constants with odd parity, e.g., 00000000 1; therefore, one
comparisoncircuit or syndrome generationcircuit is required
to detect the error. This detection method has been modified
by Bertoni et al. [3] andthe extra memorysize is reducedfrom
i256 bytes to i256 bits. Additionally, i9 b-EXORs
are introduced. The error detection of one byte, appended
with one-bit parity, requires eight b-EXORs (bit EXOR
operation) or a total of 16 8 b-EXORs for a 128-bit data
block. However, Bertoni et al.s scheme must predict the
output parity in MixColumns, therefore, the extra calcula-
tions of 16 4 b-EXORs are required in the encryption
process. In decryption, the error-detection hardware for
InvMixColumns is more complicated than in encryption.
Because the prediction of InvMixColumn is not derived in
[1], the cost is not specified in Table 1. The costs of Karri
et al.s scheme required the inversion of each operation and
it was also time-consuming. The operations in the key
expansion are similar to the four major operations of AES;
thus, the detailed comparisons of the key expansion are not
discussed. Although most operations require 16 B-EXORs to
compute the syndrome, it is possible to achieve the
computation with less B-EXORs.
7 ERROR DETECTION CAPABILITY
In Karri et al. [7], because the four operations of AES are
bijective, their error detection capability is very high. If it is
assumed that only one 128-bit error occurs during encryp-
tion or decryption, then all nonzero error patterns can be
detected in the operation-level, round-level, or algorithm-
level detection. In Bertoni et al. [1], they used the parity-
based technique and the undetectable errors do exist.
Bertoni et al. [1] did a lot of tests to obtain the results
about error detection capability and the results will be
compared to ours in Fig. 14.
All simulations and statements of our proposed schemes,
addressed here, are also under the three assumptions given
in Section 4. Three architectures, USBD, HSBD, and PbSBD,
were proposed herein; each architecture has three types of
CRC, 17. 16, 9. 8, and 5. 4 CRCs, as shown in Table 2.
Thus, nine methods were simulated. In PbBSD, the data
procedure is thoroughly protected by the i 1. i CRC;
thus, each operation has undetectable errors. However, in
USBD, the fault coverage in SubBytes is 100 percent, so the
amount of overall undetectable errors is 80 percent of that in
USBD. Similarly, in HSBD, the amount is reduced to
75 percent of that in USBD.
The simulation model is shown in Fig. 11. Each method is
simulated by 26 tests distinguished by the bit number of the
injected errors. The last test in Fig. 12, Fig. 13, and Fig. 14,
labeled as random, used error patterns with random
erroneous bit number. Each error pattern has 10
7
blocks and
thebit lengthof everyblockis 136128 8, 1442 64 8,
or 1604 32 8, respectively, for the 17. 16, 9. 8, or
5. 4 CRC. The all-one error block was considered as a
totally different state; hence, the maximum number of
erroneous bits was 135, 143, or 159 in a random test. Each
test used one data pattern of 10
7
data blocks, and every
YEN AND WU: SIMPLE ERROR DETECTION METHODS FOR HARDWARE IMPLEMENTATION OF ADVANCED ENCRYPTION STANDARD 729
TABLE 2
The Possible Combinations of Our Proposed Schemes
Fig. 11. The simulation model. Each data block has 64 ones and the
position of ones uniformly distributed in a data block. The error bits
uniformly distribute in an error block. The assignment of error blocks
uniform distributes in both rounds and operations.
Fig. 12. Percentage of undetectable errors of the 17. 16 CRC over
G12
8
.
block has 64-bit ones of normal distribution. The erroneous
rounds and erroneous operation were also randomly
chosen.
As seen in Fig. 12, all the simulated odd-bit errors were
detected. The percentage of the undetectable errors
dropped dramatically as the erroneous bit number in-
creased. When the number of erroneous bits was greater
than eight, the percentage was below 1 percent and stable.
The test using random erroneous bits is about 0.3 percent
and it was close to the theoretic value obtained in Section 4,
0.4 percent. Obviously, all the experimental results followed
the curves of ideal values.
The same data patterns used in the above tests were also
used for the 9. 8 CRC and the 5. 4 CRC; all test
conditions, except for the error patterns, were identical to
those used to test the 17. 16 CRC. The 9. 8 CRC generated
two parities for a 128-bit data block. Because the values in
the two tests, 2-bit and 4-bit erroneous bits, are too large,
they were dependently shown in Fig. 13. All odd-bit errors
were also detected. The percentage also dropped dramati-
cally when the erroneous bits increased, as shown in Fig. 13.
For the random test, the percentage is about 0.14 10
2
%,
very close to the theoretical value of 0.16 10
2
%.
In Fig. 14, the results of the 5. 4 CRC and Bertoni et al.
[1] are shown. Obviously, this percentage is very small in
contrast to the 17. 16 CRC or the 9. 8 CRC. When the
number of erroneous bits was larger than 16, the percen-
tages of undetectable errors dropped to zero. The percen-
tage in the random test was 0 percent, very close to the
theoretic value of 2.56 10
8
%. Of course, all odd-bit errors
could be detected.
Fig. 14 also shows the results in Bertoni et al. [1]. The test
models of Bertoni et al. [1] are different from ours. They
have injected multiple bit errors (between 2 to 16) at the
beginning of the round. From Fig. 14, their scheme has
better error detection than ours, when the errors are
between 2 and 6, and the cases of 8-bit errors are close.
When the number of erroneous bits is above 10, the
performance of the proposed scheme is better than that of
Bertoni et al. [1].
8 CONCLUSIONS
This work has proposed a simple, symmetric, and high-
fault-coverage error detection scheme for AES. Although
the erroneous bits are diffused in AES, this work used the
linear behavior of each operation in AES to design a
detection scheme. This scheme only uses an i 1. i CRC
to detect the errors, where i 2 f4. 8. 16g, and the parity of
the output of each operation is predicted in a simple
fashion. Even though the number of parities is two or four,
respectively, for i 8 or i 4, it is possible to use only one
8-bit register for storing the parities during hardware
implementation. This error detection may also be used in
encryption-only or decryption-only designs. Because of the
symmetry of the proposed detection scheme, the encryption
and decryption circuit can share the same error detection
hardware. The proposed schemes can be applied in the
implementation of AES against differential fault attacks and
can be easily implemented in a variety of structures, such as
8-bit, 32-bit, or 128-bit structures.
ACKNOWLEDGMENTS
This work was financially supported by the Program for
Promoting University Academic Excellence under Grant no.
EX-91-E-FA06-4-4.
REFERENCES
[1] G. Bertoni, L. Brevegelieri, I. Koren, P. Maistri, and V. Piuri, Error
Analysis and Detection Procedures for a Hardware Implementa-
tion of the Advanced Encryption Standard, IEEE Trans. Compu-
ters, vol. 52, no. 4, pp. 492-505, Apr. 2003.
[2] G. Bertoni, L. Brevegelieri, I. Koren, P. Maistri, and V. Piuri,
Detecting and Locating Faults in VLSI Implementations of the
Advanced Encryption Standards, Proc. 18th IEEE Intl Symp.
Defect and Fault Tolerance in VLSI Systems, pp. 105-113, Nov. 2003.
730 IEEE TRANSACTIONS ON COMPUTERS, VOL. 55, NO. 6, JUNE 2006
Fig. 13. Percentage of undetectable errors of the 9. 8 CRCover G12
8
.
Their percentage is 4.14 percent for 2-bit errors and 0.067 percent for
4-bit errors.
Fig. 14. Percentage of undectable errors of the 5. 4 CRC over G12
8
.
The percentage is 1.8 percent for 2-bit errors and 0.13 percent for
4-bit errors.
[3] G. Bertoni, L. Brevegelieri, I. Koren, and P. Maistri, An Efficient
Hardware-based Fault Diagnosis Scheme for AES: Performances
and Cost, Proc. 19th IEEE Intl Symp. Defect and Fault Tolerance in
VLSI Systems, pp. 130-138, Oct. 2004.
[4] E. Biham and A. Shamir, Differential Fault Analysis of Secret Key
Cryptosystems, Advances in CryptologyProc. CRYPTO 97,
pp. 513-525, 1997.
[5] P. Dusart, G. Letourneux, and O. Vivolo, Differential Fault
Analysis on A.E.S, Applied Cryptography and Network Security,
pp. 293-306, 2003.
[6] M. Feldhofer, S. Dominikus, and J. Wolkerstorfer, Strong
Authentication for RFID Systems Using the AES Algorithm,
Proc. Cryptographic Hardware and Embedded Systems (CHES 04),
pp. 357-370, 2004.
[7] R. Karri, K. Wu, P. Mishra, and Y. Kim, Concurrent Error
Detection Schemes for Fault-Based Side-Channel Cryptanalysis of
Symmetric Block Ciphers, IEEE Trans. Computer-Aided Design of
Integrated Circuits and Systems, vol. 21, no. 12, pp. 1509-1517, Dec.
2002.
[8] R. Karri, G. Kuznetsov, and M. Goessel, Parity-Based Concurrent
Error Detection of Subsititution-Permutation Network Block
Ciphers, Proc. Cryptographic Hardware and Embedded Systems
(CHES 03), pp. 113-124. 2003.
[9] S. Mangard, M. Aigner, and S. Dominikus, A Highly Regular and
Scalable AES Hardware Architecture, IEEE Trans. Computers,
vol. 52, no. 4, pp. 483-491, Apr. 2003.
[10] US Natl Inst. of Standards and Technology, Federal Information
Processing Standards Publication 197Announcing the
ADVANCED ENCRYPTION STANDARD (AES), 2001, http://
csrc.nist.gov/publications/fips/fips197/fips-197.pdf.
[11] G. Piret and J.J. Quisquater, A Differential Fault Attack
Technique against SPN Structures, with Application to the AES
and KHAZAD, Proc. Cryptographic Hardware and Embedded Systems
(CHES 03), pp. 77-88, 2003.
[12] J. Daemen and V. Rijmen, AES Proposal: Rijndael, AES
Algorithm Submission, Sept. 1999.
[13] A. Satoh, S. Morioka, K. Takano, and S. Munetoh, A Compact
Rijndael Hardware Architecture with S-Box Optimization, Proc.
Advances in Cryptology (ASIACRYPT 01), pp. 171-184, 2001.
[14] K. Wu, R. Karri, G. Kuznetsov, and M. Goessel, Low Cost
Concurrent Error Detection for the Advanced Encryption Stan-
dard, Proc. Intl Test Conf. (ITC 04), pp. 1242-1248, 2004.
Chi-Hsu Yen received the BS degree in
electrical engineering from National Central
University, Jhongli, Taiwan, in 1995 and the
MS degree in electrical engineering from Tam-
kang University, Tamsui, Taiwan, in 1997. He
received the PhD degree from the Department of
Electrical and Control Engineering of National
Chiao-Tung University, Hsinchu, Taiwan, in
2005. His research interests include crypto-
graphic algorithms and error-control coding.
Bing-Fei Wu (S89-M92-SM02) received the
BS and MS degrees in control engineering from
National Chiao Tung University (NCTU),
Hsinchu, Taiwan, in 1981 and 1983, respec-
tively, and the PhD degree in electrical engineer-
ing from the University of Southern California,
Los Angeles, in 1992. Since 1992, he has been
with the Department of Electrical Engineering
and Control Engineering, where he is currently a
professor. He has been involved in the research
of intelligent transportation systems for many years and is leading a
team to develop the first Smart Car with autonomous driving and active
safety system in Taiwan. His current research interests include vision-
based intelligent vehicle control, multimedia signal analysis, embedded
systems, and chip design. He is a senior member of the IEEE. He
founded and served as the chair of the IEEE Systems, Man, and
Cybernetics Society Taipei Chapter in Taiwan, 2003. He was the
director of the Research Group of Control Technology of Consumer
Electronics in the Automatic Control Section of the National Science
Council (NSC), Taiwan, from 1999 to 2000. As an active industry
consultant, he was also involved in the chip design and applications of
the flash memory controller and 3C consumer electronics in multimedia
systems. The research has been honored by the Ministry of Education
as the Best Industry-Academics Cooperation Research Award in 2003.
He received the Distinguished Engineering Professor Award from the
Chinese Institute of Engineers in 2002, the Outstanding Information
Technology Elite Award from the Taiwan Government in 2003, the
Golden Linux Award in 2004, the Outstanding Research Award in 2004
from NCTU, the Research Awards from NSC in the years of 1992, 1994,
1996-2000, the Golden Acer Dragon Thesis Award sponsored by the
Acer Foundation in 1998 and 2003, respectively, the First Prize Award of
the We Win (Win by Entrepreneurship and Work with Innovation &
Networking) Competition hosted by Industrial Bank of Taiwan in 2003,
and the Silver Award of Technology Innovation Competition sponsored
by the Advantech Foundation in 2003.
> For more information on this or any other computing topic,
please visit our Digital Library at www.computer.org/publications/dlib.
YEN AND WU: SIMPLE ERROR DETECTION METHODS FOR HARDWARE IMPLEMENTATION OF ADVANCED ENCRYPTION STANDARD 731