Cryptographic algorithms are often mathematically "perfect," yet they fail in the real world because they must run on physical hardware. When a chip performs a calculation, it consumes power and emits electromagnetic radiation. These physical signals—side channels—leak information about the internal data being processed.
The research direction across these papers focuses on the cat-and-mouse game between masking, a defense that splits secrets into multiple random shares, and higher-order side-channel analysis, an attack strategy that recombines those shares through statistical correlation.
The Intuition of Masking
To protect a secret value $x$, a designer can use Boolean masking. Instead of processing $x$ directly, the hardware splits it into two random shares, $s_1$ and $s_2$, such that $x = s_1 \oplus s_2$. If an attacker observes the power consumption of the chip while it processes only $s_1$, they learn nothing about $x$ because $s_1$ is just a random number.
Researchers like Amir Moradi and Axel Pöschmann have demonstrated that these defenses can be implemented even in ultra-constrained hardware, such as RFID tags, using Threshold Implementations. These implementations are specifically designed to be "first-order" resistant, meaning no single point in time leaks information about the secret.
Why First-Order Defense Fails
The core problem, as identified by Limin Guo and Kaiyan Chen, is that first-order masking only protects against attackers looking at one variable at a time. A higher-order attack looks at the joint leakage of multiple shares.
Imagine an attacker records the power consumption at two different moments: $t_1$ (when $s_1$ is processed) and $t_2$ (when $s_2$ is processed). By correlating the product or the statistical variance of these two points, the random masks "cancel out," revealing the underlying secret $x$. Weijia Wang and Zhuo Huang have shown that even modern post-quantum algorithms like HQC are vulnerable to these second-order attacks, which can bypass first-order masking with only a few thousand power traces.
Finding Needles in High-Dimensional Haystacks
For a higher-order attack to work, the attacker must know exactly which points in a power trace (which may contain millions of samples) correspond to the different shares. This is a dimensionality problem.
Recent work has turned to deep learning to automate this search. Zijia Yang and colleagues use autoencoders to reduce the dimensionality of side-channel data, making it easier to find the "Points of Interest" (PoIs). To understand what these neural networks are actually seeing, Shivam Bhasin and Stjepan Picek developed Key Guessing Occlusion (KGO). By systematically hiding parts of a trace and checking if the network can still recover the key, they can isolate the exact leakage points that higher-order attacks exploit.
Quantifying and Verifying Security
As attacks become more sophisticated—such as the 4,900× speedup achieved by Yongbin Zhou against Dilithium—designers need rigorous ways to prove their hardware is safe.
- Metrics: A. Adam Ding proposed the Information Leakage Amount (ILA), a unified metric that predicts the success rate of an attack before the hardware is even built.
- Formal Verification: Tools like QMVerif, developed by Pengfei Gao and others, use model-counting to check if a program is "perfectly masked."
The Future: Post-Quantum and Neural Hardware
The frontier of this research lies in applying these concepts to new domains. Yutaro Tanaka and Rei Ueno have moved beyond simple binary checks to multi-valued side-channel oracles, extracting multiple bits of information per trace to break Kyber. Meanwhile, Anuj Dubey is extending masking to protect machine learning hardware itself, ensuring that the weights of a neural network don't leak through the power lines of the AI chips they run on.