Introduction
In the evolving landscape of cryptanalysis, deep learning (DL) has emerged as a potent tool for identifying vulnerabilities in classic ciphers. While "neural-differential" attacks have achieved notable success, modernizing "linear attacks" with AI has remained a significant challenge. Linear cryptanalysis, introduced by Mitsuru Matsui in 1993, relies on finding linear approximations—equations relating plaintext, ciphertext, and key bits with a probability slightly deviating from 1/2. The primary hurdle for DL models is that this statistical bias is often too subtle to be detected from individual plain-ciphertext pairs.
A recent study, "Neural-Linear Attack Based on Distribution Data and Its Application on DES," introduces a novel preprocessing technique called "distribution data." By shifting the focus from raw ciphertexts to sequences of linear approximation results, the authors significantly enhance the efficiency of neural-linear attacks on the Data Encryption Standard (DES).
Methodology: The Power of Distribution Data
The core innovation of this research is the transition from structural learning to statistical learning. Previous neural-linear distinguishers attempted to learn from raw plain-ciphertext pairs, which often failed because a single pair lacks sufficient information to represent a statistical distribution.
Preprocessing and Vectorization
Instead of treating one pair as a single training sample, the authors group $N$ pairs together. For each pair, they compute the XOR sum of specific bits according to a linear approximate equation. These $N$ results are then concatenated into an $N$-bit binary string. This string effectively serves as a sample from a binomial distribution, making the underlying statistical bias "visible" to the neural network.
Model Architecture
The researchers utilized a Residual Network (ResNet) architecture tailored for this distribution-based input. A critical modification is the inclusion of a "Reshape" layer that transforms the $N$-bit distribution strings into matrices (e.g., $N/8 \times 8$), allowing standard convolutional layers to process the statistical patterns effectively.
Experimental Setup and Datasets
The attack was evaluated on round-reduced versions of DES (3, 4, 5, and 6 rounds). The datasets consisted of:
- Training Sets: $10^4$ to $10^5$ distribution samples.
- Validation Sets: Approximately $10^3$ samples.
- Input Lengths ($N$): Various lengths (16, 32, 64, 256, 512, 1024) were tested to quantify the relationship between data volume per sample and model performance.
Key Insights: The "N" Factor
A pivotal discovery was the direct correlation between the length of the distribution string ($N$) and the accuracy of the neural distinguisher. As $N$ increases, the statistical signal becomes clearer, allowing the ResNet to distinguish biased distributions from random noise with higher confidence.
On 4-round DES, increasing $N$ from 64 to 512 boosted accuracy from 81.5% to an impressive 99.6%. Furthermore, the researchers found that a distinguisher trained on a simple one-bit linear equation could be applied to complex multi-bit key recovery without retraining, as the fundamental task of identifying biased binomial distributions remains constant.
Results and Key Recovery
The proposed method represents a substantial improvement over previous state-of-the-art neural-linear attacks:
- Distinguisher Accuracy: On 4-round DES, the accuracy reached 99.6%, a massive leap from the 53.75% reported in earlier literature.
- One-Bit Recovery: For 5-round DES, a 91.15% success rate was achieved using only 2048 plain-ciphertext pairs, outperforming traditional theoretical benchmarks.
- Breaking 6 Rounds: The attack successfully reached 6-round DES with a 60.6% success rate, a milestone that previous neural methods struggled to achieve reliably.
- Practical Multi-Bit Recovery: New algorithms were introduced to rank subkey candidates. The correct key bits were consistently ranked within a small searchable range (average rank of 30), demonstrating the practical viability of the attack.
Conclusion and Future Work
By rethinking the input representation as distribution data, this research bridges the gap between traditional statistical cryptanalysis and modern deep learning. While the attack on 6-round DES is a significant achievement, scaling to the full 16 rounds remains a challenge due to the exponential decay of linear bias. Future work will focus on refining key ranking algorithms and exploring multiclass classification to identify multiple key bits simultaneously, further optimizing the efficiency of neural-linear attacks.
Summary updated