Published in arXiv preprint, 2024, 2024
This paper proposes Mutual Information Preserving Pruning (MIPP), a structured activation-based pruning method that selects nodes to conserve mutual information shared between activations of adjacent layers.
Recommended citation: Westphal, C., Hailes, S., & Musolesi, M. (2024). Mutual Information Preserving Neural Network Pruning. arXiv preprint arXiv:2411.00147.
Download Paper
Published in AISTATS 2025, 2025
This paper introduces PIDF, a novel approach that simultaneously addresses data interpretability and feature selection by evaluating each feature using three metrics: mutual information, synergistic contribution, and redundancy.
Recommended citation: Westphal, C., Hailes, S., & Musolesi, M. (2024). Partial Information Decomposition for Data Interpretability and Feature Selection. arXiv preprint arXiv:2405.19212.
Download Paper
Published in SIGKDD 2025, 2025
This paper presents FSNID, a novel information-theoretic method that facilitates the exclusion of non-informative features when detecting network intrusions, selecting a significantly reduced feature set while maintaining performance.
Recommended citation: Westphal, C., Hailes, S., & Musolesi, M. (2025). Feature Selection for Network Intrusion Detection. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining (pp. 1599-1610).
Download Paper
Published in arXiv preprint, 2025, 2025
This paper introduces a Generalized Information Bottleneck framework that reconceptualizes the original Information Bottleneck principle through synergy—information accessible only via joint feature processing.
Recommended citation: Westphal, C., Hailes, S., & Musolesi, M. (2025). A Generalized Information Bottleneck Theory of Deep Learning. arXiv preprint arXiv:2509.26327.
Download Paper
Published in TMLR 2026, 2026
This paper introduces the Transfer Entropy Redundancy Criterion (TERC), an information-theoretic approach for identifying optimal state variables in RL that provably excludes variables with no effect on agent performance. Accepted to TMLR.
Recommended citation: Westphal, C., Hailes, S., & Musolesi, M. (2026). Information-theoretic State Variable Selection for Reinforcement Learning. Transactions on Machine Learning Research (TMLR).
Download Paper
Published in arXiv preprint, 2026, 2026
Shows that activation-based steganography detection in LLMs can be evaded via adversarial fine-tuning, and proposes a data-level intervention using recontextualization datasets that restores detectability.
Recommended citation: Westphal, C., Douglas, T., Navaie, K., Pimentel, T., & Rosas, F. E. (2026). Now You (Still) See Me: Detecting Evasive Steganographic Payloads in LLMs. arXiv preprint arXiv:2606.09411.
Download Paper
Published in ICML 2026, 2026
This paper introduces geometry-based steganography in fine-tuned LLMs using embedding-space-derived mappings, and proposes mechanistic-interpretability-based detection via linear probes on model activations. Accepted to ICML 2026.
Recommended citation: Westphal, C., Navaie, K., & Rosas, F. E. (2026). Hide and Seek in Embedding Space: Geometry-based Steganography and Detection in Large Language Models. To appear in Proceedings of the International Conference on Machine Learning (ICML).
Download Paper
Published:
I will present my work on Partial Information Decomposition (PIDF) as part of the opening session of AiStats 2025. This will cover novel methods for quantifying redundant, unique, and synergistic information in high-dimensional datasets, with applications to interpretable machine learning and feature selection.
Published:
I will present my research on information-theoretic feature selection for network intrusion detection systems. This work introduces novel methods to identify minimal yet maximally discriminative features in cybersecurity datasets, significantly reducing computational overhead while maintaining state-of-the-art detection accuracy. The talk will include: