A Quantization-Aware Training Based Lightweight Method for Neural Distinguishers

2026-03-06

AI Summary

This work addresses the high computational overhead of traditional neural distinguishers, which rely on 32-bit multiplications and are thus ill-suited for lightweight cryptanalytic applications. The authors propose a quantization-aware trained lightweight neural distinguisher that introduces learnable-step-size quantization to this domain for the first time, compressing weights to 1.58 bits. By replacing convolutional multiplications with Boolean operations and reformulating ReLU as a comparison-based indicator function, they construct an efficient architecture comprising only Boolean operations, additions, and indicator functions. Experiments demonstrate the complete elimination of 32-bit multiplications, reducing total computational cost to 13.9% of that in Gohr’s model with only a 2.87% drop in overall accuracy. Notably, replacing the first layer’s 128 one-by-one convolutions with four 16-bit Boolean operations incurs merely a 0.3% precision loss.

Technology Category

• Machine Learning: Hardware-aware ML
• Computer Vision: Adversarial Attacks & Robustness
• Search and Optimization: Learning to Search

Application Category

• Responsible Web: Measurement, analysis, and circumvention of Web censorship
• Security and Privacy: Large-scale security measurements
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics

Abstract

In 2019, Gohr pioneered the application of deep neural networks to differential cryptanalysis, developing DNN-based neural distinguisher classifiers to analyze the SPECK lightweight block cipher. Unlike traditional differential analysis, which relies on Boolean operations on 0-1 sequences, neural distinguishers extract continuous features, introducing 32-bit multiplications operations that increase complexity and potential redundancy. This study proposes a lightweight neural distinguisher based on quantization-aware training. Leveraging learnable step-size quantization, the model's weights are quantized to 1.58 bits, enabling the replacement of all convolutional multiplication operations with Boolean logic. Additionally, the ReLU activation function is reimplemented as a comparison-based indicator function. This transforms the original 32-bit multiplication-dependent architecture into a lightweight structure composed solely of Boolean operations, additions, and indicator functions. Experimental results confirm significant computational complexity reduction. Owing to a high proportion of zero-valued weights, the total operations amount to just 13.9% of Gohr's model. Critically, the most costly 32-bit multiplications are eliminated, with classification accuracy dropping by only 2.87%. When applied exclusively to the initial convolutional layer, the 128 1-by-1 convolutions are replaced with 4 Boolean operations on 16-bit sequences, incurring a negligible 0.3% accuracy loss.

Problem

• neural distinguisher
• differential cryptanalysis
• lightweight
• quantization
• computational complexity

Innovation

• quantization-aware training
• neural distinguisher
• lightweight cryptography
• Boolean operations
• learnable step-size quantization