International Association
for Cryptologic Research

Side-channel attacks on elliptic curve cryptography (ECC) often assume a white-box attacker who has detailed knowledge of the implementation choices taken by the target implementation. Due to the complex and layered nature of ECC, there are many choices that a developer makes to obtain a functional and interoperable implementation. These include the curve model, coordinate system, addition formulas, and the scalar multiplier, or lower-level details such as the finite-field multiplication algorithm. This creates a gap between the attack requirements and a real-world attacker that often only has black-box access to the target – i.e., has no access to the source code nor knowledge of specific implementation choices made. Yet, when the gap is closed, even real-world implementations of ECC succumb to side-channel attacks, as evidenced by attacks such as TPM-Fail, Minerva, the Side Journey to Titan, or TPMScan [MSE+20; JSS+20; RLM+21; SDB+24].We study this gap by first analyzing open-source ECC libraries for insight into realworld implementation choices. We then examine the space of all ECC implementations combinatorially. Finally, we present a set of novel methods for automated reverse engineering of black-box ECC implementations and release a documented and usable open-source toolkit for side-channel analysis of ECC called pyecsca.Our methods turn attacks around: instead of attempting to recover the private key, they attempt to recover the implementation configuration given control over the private and public inputs. We evaluate them on two simulation levels and study the effect of noise on their performance. Our methods are able to 1) reverse-engineer the scalar multiplication algorithm completely and 2) infer significant information about the coordinate system and addition formulas used in a target implementation. Furthermore, they can bypass coordinate and curve randomization countermeasures.

This paper describes an ECC implementation computing the X25519 keyexchange protocol on the Arm Cortex-M4 microcontroller. For providing protections against various side-channel and fault attacks we first review known attacks and countermeasures, then we provide software implementations that come with extensive mitigations, and finally we present a preliminary side-channel evaluation. To our best knowledge, this is the first public software claiming affordable protection against multiple classes of attacks that are motivated by distinct real-world application scenarios. We distinguish between X25519 with ephemeral keys and X25519 with static keys and show that the overhead to our baseline unprotected implementation is about 37% and 243%, respectively. While this might seem to be a high price to pay for security, we also show that even our (most protected) static implementation is at least as efficient as widely-deployed ECC cryptographic libraries, which offer much less protection.

The adoption of deep neural networks for profiled side-channel attacks provides powerful options for leakage detection and key retrieval of secure products. When training a neural network for side-channel analysis, it is expected that the trained model can implement an approximation function that can detect leaking side-channel samples and, at the same time, be insensible to noisy (or non-leaking) samples. This outlines a generalization situation where the model can identify the main representations learned from the training set in a separate test set.This paper discusses how output class probabilities represent a strong metric when conducting the side-channel analysis. Further, we observe that these output probabilities are sensitive to small changes, like selecting specific test traces or weight initialization for a neural network. Next, we discuss the hyperparameter tuning, where one commonly uses only a single out of dozens of trained models, where each of those models will result in different output probabilities. We show how ensembles of machine learning models based on averaged class probabilities can improve generalization. Our results emphasize that ensembles increase a profiled side-channel attack’s performance and reduce the variance of results stemming from different hyperparameters, regardless of the selected dataset or leakage model.

Fault Injection (FI) attacks have become a practical threat to modern cryptographic implementations. Such attacks have recently focused more on exploitation of implementation-centric and device-specific properties of the faults. In this paper, we consider the parallel between SCA attacks and FI attacks; specifically, that many FI attacks rely on the data-dependency of activation and propagation of a fault, and SCA attacks similarly rely on data-dependent power usage. In fact, these are so closely related that we show that existing SCA attacks can be directly applied in a purely FI setting, by translating power FI results to generate FI ‘probability traces’ as an analogue of power traces. We impose only the requirements of the equivalent SCA attack (e.g., knowledge of the input plaintext for CPA on the first round), along with a way to observe the status of the target (whether or not it has failed and been “muted” after a fault). We also analyse existing attacks such as Fault Template Analysis in the light of this parallel, and discuss the limitations of our methodology. To demonstrate that our attacks are practical, we first show that SPA can be used to recover RSA private exponents using FI attacks. Subsequently, we show the generic nature of our attacks by performing DPA on AES after applying FI attacks to several different targets (with AVR, 32-bit ARM and RISC-V CPUs), using different software on each target, and do so with a low-cost (i.e., less than $50) power fault injection setup. We call this technique Fault Correlation Analysis (FCA), since we perform CPA on fault probability traces. To show that this technique is not limited to software, we also present FCA results against the hardware AES engine supported by one of our targets. Our results show that even without access to the ciphertext (e.g., where an FI redundancy countermeasure is in place, or where ciphertext is simply not exposed to an attacker in any circumstance) and in the presence of light jitter, FCA attacks can successfully recover keys on each of these targets.

To mitigate side-channel attacks, real-world implementations of public-key cryptosystems adopt state-of-the-art countermeasures based on randomization of the private or ephemeral keys. Usually, for each private key operation, a “scalar blinding” is performed using 32 or 64 randomly generated bits. Nevertheless, horizontal attacks based on a single trace still pose serious threats to protected ECC or RSA implementations. If the secrets learned through a single-trace attack contain too many wrong (or noisy) bits, the cryptanalysis methods for recovering remaining bits become impractical due to time and computational constraints. This paper proposes a deep learning-based framework to iteratively correct partially correct private keys resulting from a clustering-based horizontal attack. By testing the trained network on scalar multiplication (or exponentiation) traces, we demonstrate that a deep neural network can significantly reduce the number of wrong bits from randomized scalars (or exponents).When a simple horizontal attack can recover around 52% of attacked multiple private key bits, the proposed iterative framework improves the private key accuracy to above 90% on average and to 100% for at least one of the attacked keys. Our attack model remains fully unsupervised and excludes the need to know where the error or noisy bits are located in each separate randomized private key.