In , Biham and Shamir announce an attack on DES based on 200 ciphertexts in which one-bit errors have been induced by environmental stress. Here we show an attack that requires less than ten ciphertexts. Furthermore, our attack is practical in that it uses a fault model that has been implemented in attacks on real smartcards.
In , Biham and Shamir show how their method can be extended to reverse engineer algorithms whose structure is unknown. Our attack can also be extended to such cases and is more efficient there too. In , Boneh, De Millo and Lipton discuss how such techniques can be used to attack RSA. Again, their attack is theoretical only. We show how to do it in practice.
A recent research announcement by Biham and Shamir shows that if DES is implemented in a tamper-resistant package, and this package has the property that by applying ionising radiation (or some other environmental stress) we can cause random one-bit errors in the round keys, then we can break DES. If we can get a series of ciphertexts, each generated from the same plaintext but with a different one-bit random round key error, then we will need about 200 faulty ciphertexts to recover the key. In a further announcement , they show how on a similar fault model, the structure of unknown Feistel ciphers can be deduced from an adequate number of faulty ciphertexts. In each case, the critical observation is that errors that occur in the last few round of the cipher leak information about the key, or algorithm structure, respectively.
This work is inspired by a paper of Boneh, DeMillo and Lipton  who assume (as Biham and Shamir do) that one-bit errors can be caused by radiation or other environmental stresses. That paper goes on to show that with this fault model, RSA can be attacked.
These results have been widely publicised in the press. A frequently voiced criticism is that the attacks are purely theoretical: no-one has demonstrated that single bit errors can actually be induced in a DES key schedule or an RSA computation in any fielded device. In fact, most smartcards hold keys in EEPROM which also contains much or all of the device's application software. Thus errors induced by ionising radiation would be much more likely to corrupt software, thus leading either to a system crash or to uninformative wrong answers.
We show here that much faster, and completely practical, attacks are possible. The trick is to induce small changes in the code rather than trying to cause them in keys or other data.
The Improved Attack Methodology
In a note posted to relevant Internet newsgroups on the 7th November, one of us pointed out that using clock and power glitches gives a practical way of implementing the Lenstra variant of the Boneh attack. In this announcement, we will expand on the threat model, and also show how attacks using clock and power glitches can give attacks on DES that require many fewer ciphertexts - less than ten rather than the 200 or so previously required.
In a paper on tamper resistance due to be published next week, we describe a number of techniques for attacking smartcards and other security processors . This paper was written some time ago (the first results were presented at the Isaac Newton Institute, Cambridge, in June) but has been withheld by agreement with the manufacturer of one of the security processors we have attacked, so that banking industry clients had some time to take suitable precautions. Some of the attacks we describe in this paper are new, while others are already known in various small communities (such as hackers and chip testers) and are included for the benefit of the wider crypto and security communities.
One of the latter type is that smartcards and other security processors can often be attacked using clock and power glitches. The application of a clock pulse that is much faster than normal, or of a transient in the power supply, can often cause faulty behaviour in a microprocessor, under which the program counter is incremented but the current instruction is executed either incorrectly or not at all. A standard version of this attack is to replace a single 5MHz clock pulse to a smartcard with four 20MHz pulses.
We do not claim to have invented this attack; it appears to have originated in the pay-TV hacking community, which has known about it for at least a year. In that context, it has not been used for attacks on cryptographic algorithms, but in order to cause output loops to run for longer than the card's programmer intended, thus dumping key material to the output port.
The glitch attack is described more fully in our tampering article, which will appear at next week's Usenix Electronic Commerce Workshop .
In this note, we point out that attacks based on faulty instructions are not only proven practical, unlike the as yet undemonstrated fault model of random single bit errors induced by radiation. They also provide a much more powerful attack on many cryptographic algorithms. This holds both when we are seeking to recover a key for a known algorithm such as DES, and when we are trying to reverse engineer an unknown algorithm that has been provided in a smartcard or other tamper resistant processor.
A simplified version of the Boneh-DeMillo-Lipton attack, due to Lenstra, is quoted in : if a smart card computes an RSA signature S on a message M modulo n = pq by computing it modulo p and q separately and then combining them using the Chinese Remainder Theorem, and if an error an be induced in (say) the latter computation, then we can factor n at once as p = gcd(n,S^e-M) where e is the public exponent.
This is absolutely ideal for a glitch attack. As the card spends most of its time calculating the signature mod p and mod q, and almost any glitch that affects the output will do, we do not have to be at all selective about where in the instruction sequence the glitch is applied. Since only a single signature is needed, the attack can be performed online: a Mafia eftpos terminal applies the glitch, factors the modulus, calculates what the correct signature should be, and sends this on to the bank.
Thus the Mafia can harvest RSA secret keys without the customer or his bank noticing anything untoward about the transaction he did at their shop. Given that implementers of the new EMV electronic purse system propose to have only 10,000 different RSA secret keys per issuing bank, the Mafia will soon be able to forge cards for a substantial proportion of the user population.
When we can cause an instruction of our choice to fail, then attacking DES is simple. We remove one of the xor operations that are used to combine the round keys with the inputs to the S-boxes from the last two rounds of the cipher, and repeat this for each of these key bytes in turn. The erroneous ciphertext outputs that we receive as a result of this attack will each differ from the genuine ciphertext in the output of usually two, and sometimes three, S-boxes. Using the techniques of differential cryptanalysis, we obtain about five bits of information about the eight keybits that were not xor'ed as a result of the induced fault. So, for example, eight faulty ciphertexts should give us about 40 bits of the key, leaving a trivial keysearch.
Thus DES can be attacked with about one correct and eight faulty ciphertexts. But how realistic is it to assume that we will be able to target particular instructions?
In most smartcards, the manufacturer supplies a number of routines in ROM. Though sometimes presented as an `operating system', the ROM code is more of a library or toolkit that enables application developers to manage communications and other facilities. Its routines usually include the DES algorithm (or a proprietary algorithm such as Telepass), and by buying the manufacturer's smartcard development toolkit (for typically a few thousand dollars) an attacker can get full documentation plus real specimens for testing. In this case, individual DES instructions can be targeted.
When confronted with an unfamiliar implementation, we may have to experiment somewhat (we have to do this anyway with each card in order to find the correct glitch parameters ). However the search space is relatively small, and on looking at a few DES implementations it becomes clear that we can usually recognise the effects of removing a single instruction from either of the last two rounds. (In fact, many of these instructions yield almost as much information when removed from the implementation as the key xor instructions do.) We will discuss this at greater length in a later paper.
The ROM Overwrite Attack
Where the implementation is familiar, there is yet another way to extract keys from the card - the ROM overwrite attack.
Single bits in a ROM can be overwritten using a laser cutter, and where the DES implementation is well known, we can find one bit (or a small number of bits) with the property that changing it will enable the key to be extracted easily. The details will depend on the implementation but we might well be able, for example, to make a jump instruction unconditional and thus reduce the number of rounds in the cipher to one or two. Where the algorithm is kept in EEPROM, we can use two microprobing needles to set or reset the target bit .
Where we have incomplete information on the implementation, ROM overwriting attacks can be used in other ways. For example, if the DES S-boxes in ROM, we can identify them using an optical microscope and use our laser cutter to make all their bits equal. This turns DES into a linear transfromation over GF(2), and we can extract the key from a single plaintext/ciphertext pair.
Although ROM overwrite (unlike the other attacks suggested in this paper) involves access to the chip surface, it can be carried out using tools that are relatively cheap and widely available. So it may be used by attackers who do not have access to the expensive semiconductor test equipment that professional pirates use to extract keys directly from smartcards .
Returning to the non-invasive attack model, we can always apply clock and power glitches until simple statistical tests suddenly show a high dependency between the input and output of the encryption function, indicating that we have succeeded in reducing the number of rounds. This may be practical even where the implementation details are unknown.
Reverse Engineering an Unknown Block Cipher
In , Biham and Shamir discuss how to identify the structure of an unknown block cipher in a tamper resistant package (e.g., Skipjack) using one-bit random errors. As before, they identify faults that affected only the last round or rounds; this can be done by looking for ciphertexts at a low Hamming distance from each other. They then identify which output bits correspond to the left and right halves, and next look at which bits in the left half are affected by one bit changes in the last-but-one right half. In the case of a cipher such as DES with S boxes, the structure will quickly become clear and with enough ciphertexts the values of the S-boxes can be reconstructed. They report that with 500 ciphertexts the gross structure can be recovered, and with about 10,000 the S-box entries themselves can be found.
Our technique of causing faults in instructions rather than in data bits is more effective here too. We can attack the last instruction, then the second last instruction, and so on.
We will give detailed estimates for DES in the final paper. Let us now consider an actual classified algorithm. `Red Pike' was designed by GCHQ for encrypting UK government traffic classified up to `Restricted', and the Department of Health wishes to use it to encrypt medical records. The British Medical Association, advised by one of us (Anderson) instead recommended that an algorithm be chosen that had been in the open literature for at least two years and had withstood attempts to find shortcuts (triple-DES, Blowfish, SAFER K-128, WAKE,...).
In order to try and persuade the BMA that Red Pike was sound, the government commissioned a study of it by four academics . This study states that Red Pike `uses the same basic operations as RC5' (p 4) in that the principal operations are add, exclusive or, and left shift. It `has no look-up tables, virtually no key schedule and requires only five lines of code' (p 4). Other hints include that `the influence of each key bit quickly cascades' (p 10) and `each encryption involves of the order of 100 operations' (p 19).
We can thus estimate the effort of reverse engineering Red Pike from a tamper resistant hardware implementation by considering the effort needed to mount a similar attack on RC5.
Removing the last operation - the addition of key material - yields an output in which the right hand side is different (it is (B xor A) shl A). This suggests, correctly, that the cipher is a balanced Feistel network without a final permutation. Removing the next operation - the shift - makes clear that it was a 32 bit circular shift but without revealing how it was parametrised. Removing the next operation - the xor - is transparent, and the next - the addition of key material in the previous round - yields an output with the values A and B in the above expression. It thus makes the full structure of the data-dependent rotation clear. The attacker can now guess that the algorithm is defined by
A = ((A xor B) shl B) op key B = ((B xor A) shl A) op key
Reverse engineering RC5's rather complex key schedule (and deducing that `op' is actually +) would require single-stepping back through it separately; but once we know that `op' is +, we can find the round key bits directly by working back through the rounds of encryption.
So, apart from its key schedule, RC5 may be about the worst possible algorithm choice for secret-algorithm hardware applications, where some implementations may be vulnerable to glitch attacks. If Red Pike is similar but with a simpler key schedule, then it could be more vulnerable still. However, since the government plans to make Red Pike available eventually in software, this is not a direct criticism of the design or choice of that algorithm.
It does all suggest, though, that secret-hardware algorithms should be more complex; large S-boxes kept in EEPROM (that is separate from the program memory) may be a sensible way of pushing up the cost of an attack. Other protective measures that prudent designers would consider include error detection, multiple encryption with voting, and designing the key schedule so that the key material from a small number of rounds is not enough for a break.
We have improved on the Differential Fault Analysis of Biham and Shamir. Rather than needing about 200 faulty ciphertexts to recover a DES key, we need less than ten. We can factor RSA moduli with a single faulty ciphertext. We can also reverse engineer completely unknown algorithms; this appears to be faster than Biham and Shamir's approach in the case of DES, and is particularly easy with algorithms that have a compact software implementation such as RC5.
Finally, our attacks - unlike those of Biham, Shamir, Boneh, DeMillo and Lipton - use a realistic fault model, which has actually been implemented and can be used against fielded systems.
Mike Roe pointed out that the glitch attack on RSA can be done in real time by a Mafia owned eftpos terminal.
 ``A New Cryptanalytic Attack on DES'', E Biham, A Shamir, preprint, 18/10/96
 ``Differential Fault Analysis: Identifying the Structure of Unknown Ciphers Sealed in Tamper-Proof Devices'', E Biham, A Shamir, preprint, 10/11/96
 ``On the Importance of Checking Computations'' D Boneh, RA DeMillo, RJ Lipton, preprint
 ``A practical variant of the Bellcore attack'', RJ Anderson, posted to sci.crypt as message ID <firstname.lastname@example.org>, 7/11/96
 ``Tamper Resistance - A Cautionary Note'', RJ Anderson, MG Kuhn, to appear in Usenix Electronic Commerce workshop, 19/11/96
 ``Hardwaresicherheit von Mikrochips in Chipkarten'', Osman Kocar, Datenschutz und Datensicherheit v 20 no 7 (July 96) pp 421--424
 ``Red Pike --- An Assessment'', C Mitchell, S Murphy, F Piper, P Wild, Codes and Ciphers Ltd 2/10/96