Low-Cost Setup for Localized Semi-invasive Optical Fault Injection Attacks

Abstract

Localized semi-invasive optical fault attacks are nowadays considered to be out of reach for attackers with a limited budget. For this reason, they typically receive lower attention and priority during the security analysis of low-cost devices. Indeed, an optical fault injection setup typically requires expensive equipment which includes at least a laser station, a microscope, and a programmable X-Y table, all of which can quickly add up to several thousand euros. Additionally, a careful handling of toxic chemicals in a protected environment is required to decapsulate the chips under test and gain direct access to the die surface. In this work, we present a low-cost fault injection setup which is capable of producing localized faults in modern 8-bit and 32-bit microcontrollers, does not require handling hazardous substances or wearing protective eyeware, and would set back an attacker only a couple hundred euros. Finally, we show that the type of faults which are obtained from such a low-cost setup can be exploited to successfully attack real-world cryptographic implementations, such that of the NSA’s Speck lightweight block cipher.

Appendix: Differential Fault Analysis of Speck

Speck is a family of block ciphers variable by different block and key sizes. The round function R(x, y) of Speck has a Feistel-like structure and is described by the following equation:

$$\begin{aligned} R(x,y) := ((x\ggg \alpha + y) \oplus k, y \lll \beta \oplus (x \ggg \alpha +y) \oplus k), \end{aligned}$$

where $\oplus$ denotes a bitwise XOR, $+$ denotes an addition modulo $2^{n}$, $\ggg \alpha$ denotes the right circular shift with $\alpha$ bits, $\lll \beta$ denotes the left circular shift with $\beta$ bits, x and y are the input n-bit words, and k is the round key.

The last round key $k^{T-1}$ can be recovered by injecting random faults in the word $y^{T-1}$ as proposed by Huo et al. in [9]. The fault propagates through the last round and the pairs of correct and faulty ciphertexts are collected. Then, a system of non linear equations on $\mathbb {F}_2$ is constructed as a set of Differential Equations of Additions (DEAs). Finally, the system of DEAs is solved using a computer algebra system with the aid of Gröbner bases. According to [9], 5–8 pairs are needed on average to solve the system of DEAs, independently of the block size n.