Applications of machine learning to cryptanalysis, and increased privacy preservation through homomorphic encryption
Loading...
Files
Full Text E-thesis
Date
2026-06-01
Authors
Ebrahimi Moghaddam, Amirhossein
Journal Title
Journal ISSN
Volume Title
Publisher
University College Cork
Published Version
Abstract
This thesis investigates the two-way relationship between machine learning and cryptography in practical security settings. It examines how machine learning can support cryptanalysis and security evaluation, and how cryptographic techniques can protect sensitive data processed by machine-learning systems. Across both areas, the central concern is computational efficiency: a method must provide useful security or privacy benefits while remaining practical to train, evaluate, or deploy.
The first part focuses on machine-learning-assisted cryptanalysis of symmetric block ciphers. It introduces a Partial Differential machine-learning distinguisher that uses a selected subset of output-difference bits instead of the full block state, together with an empirical method for measuring bit effectiveness. Experiments on reduced-round SPECK32/64 show that compact feature sets can retain useful distinguishing capability while reducing computational cost. The thesis then investigates deep-learning-based rotational-XOR distinguishers for the Simon and Simeck families of AND-RX block ciphers. Evolutionary optimisation is used to explore input differences, rotation offsets, and cipher configurations. The results provide evidence that design choices, including rotation parameters and key schedules, influence RX-difference propagation and distinguishability.
The second part considers the complementary problem of protecting privacy during machine-learning inference. It presents a privacy-preserving sentiment-analysis pipeline based on the CKKS homomorphic-encryption scheme. The proposed client-server architecture combines a client-side attention mechanism with an encrypted prediction component designed to operate within ciphertext-level computational constraints. Evaluation using the IMDb dataset demonstrates the trade-off between predictive performance and the latency, memory use, and computational overhead introduced by encrypted inference.
Overall, the thesis shows that machine learning and cryptography can serve complementary roles in practical security systems. Machine learning can expose and evaluate statistical structure in cryptographic constructions, while homomorphic encryption can enable useful machine-learning services without directly revealing sensitive user inputs.
Description
Keywords
Machine learning , Cryptanalysis , Differential cryptanalysis , Rotational-XOR cryptanalysis , Homomorphic encryption , Privacy-preserving machine learning
Citation
Ebrahimi Moghaddam, A. 2026. Applications of machine learning to cryptanalysis, and increased privacy preservation through homomorphic encryption. PhD Thesis, University College Cork.