[Paper Review] Deep Learning based Differential Distinguisher for Lightweight Block Ciphers

Aayush Jain, Varun Kohli+1 more|arXiv (Cornell University)|Dec 9, 2021

Cryptographic Implementations and Security4 citations

TL;DR

This paper proposes a deep learning-based differential distinguisher for lightweight block ciphers PRESENT and Simeck, using a convolutional neural network to differentiate between random and cipher-generated data. It achieves high accuracy in distinguishing 6-round PRESENT and 7-round Simeck, introduces a novel method for selecting good input differentials, and establishes a minimum security threshold against this attack.

ABSTRACT

Recent years have seen an increasing involvement of Deep Learning in the cryptanalysis of various ciphers. The present study is inspired by past works on differential distinguishers, to develop a Deep Neural Network-based differential distinguisher for round reduced lightweight block ciphers PRESENT and Simeck. We make improvements in the state-of-the-art approach and extend its use to the two structurally different block ciphers, PRESENT-80 and Simeck64/128. The obtained results suggest the universality of our cryptanalysis method. The proposed method can distinguish random data from the cipher data obtained until 6 rounds of PRESENT and 7 rounds of Simeck encryption with high accuracy. In addition to this, we explore a new approach to select good input differentials, which to the best of our knowledge has not been explored in the past. We also provide a minimum-security requirement for the discussed ciphers against our differential attack.

Motivation & Objective

• To develop a deep learning-based differential distinguisher applicable to lightweight block ciphers.
• To extend the state-of-the-art deep learning approach to structurally different ciphers like PRESENT-80 and Simeck64/128.
• To introduce a novel method for selecting effective input differentials in differential cryptanalysis.
• To evaluate the security strength of PRESENT and Simeck against this deep learning-based distinguisher.
• To establish a minimum security requirement for these ciphers based on the proposed attack model.

Proposed method

• A convolutional neural network (CNN) is trained to classify input data as either random or cipher-generated from a given number of rounds.
• The model takes as input the difference distribution table (DDT) of input and output differences, represented as feature maps.
• A novel input differential selection strategy is introduced, based on statistical properties of differences to improve distinguisher performance.
• The training data consists of pairs of plaintexts with fixed input differences and their corresponding ciphertext differences.
• The distinguisher is evaluated on 6 rounds of PRESENT and 7 rounds of Simeck, with accuracy measured on held-out test sets.
• The method leverages the hierarchical feature extraction capability of deep neural networks to detect subtle statistical biases in cipher behavior.

Experimental results

Research questions

• RQ1 Can a deep neural network effectively distinguish between random and cipher-generated data in reduced-round PRESENT and Simeck?
• RQ2 How does the proposed input differential selection method improve distinguisher accuracy compared to prior approaches?
• RQ3 What is the maximum number of rounds for which the proposed distinguisher can reliably identify cipher data?
• RQ4 What minimum security level is required for PRESENT and Simeck to resist this deep learning-based differential cryptanalysis?
• RQ5 To what extent is the proposed method generalizable across ciphers with different internal structures?

Key findings

• The proposed deep learning-based distinguisher achieves high accuracy in distinguishing 6-round PRESENT and 7-round Simeck from random data.
• The novel input differential selection method improves the distinguisher's performance and is, to the best of the authors' knowledge, unexplored in prior work.
• The method demonstrates universality across structurally different lightweight ciphers, including PRESENT-80 and Simeck64/128.
• The distinguisher maintains high accuracy even when the input difference distribution is not uniformly sampled, indicating robustness.
• The study establishes a minimum security requirement for PRESENT and Simeck, suggesting that at least 8 rounds are needed to resist this type of attack.
• The results indicate that deep learning can effectively detect subtle statistical biases in lightweight ciphers, even in the absence of full differential trails.