Comprehensive and Reliable Feature Attribution for Diverse Modalities and Models via Frequency-Domain Insights

2026-06-06 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Explainable AI, Federated Learning · Depth: Expert, extended

Summary

FreqX is a novel interpretability method for Personalized Federated Learning (PFL), addressing challenges like Non-IID data, heterogeneous devices, fairness, and unclear client contributions that current methods fail to overcome due to high cost, privacy, and lack of detailed information. Inspired by Signal Processing and Information Theory, FreqX provides both attribution and concept information. Experiments demonstrate FreqX runs at least 10 times faster than concept-based baselines, can be transformed into a feature attribution method, and effectively extracts concept information. It also shows potential for aggregating into global explanations and evaluating client contributions in federated learning, validated across four tabular datasets (Diabetes, Phishing, BankMarketing, SpamBase) using a 3-layer MLP.

Key takeaway

For AI Scientists and Machine Learning Engineers developing Personalized Federated Learning (PFL) systems, FreqX offers a crucial interpretability solution. You can use FreqX to efficiently understand model decisions, extract both feature attribution and concept information, and accurately assess client contributions without compromising data privacy or incurring high computational costs. This enables you to design more explainable and economical PFL algorithms, improving fairness and resource allocation in heterogeneous environments. Consider integrating FreqX to enhance transparency and trust in your federated models.

Key insights

FreqX offers fast, privacy-preserving, and detailed interpretability for Personalized Federated Learning through frequency-domain insights.

Principles

• Neural networks learn frequency information.
• Forward pass increases signal-to-noise ratio.
• Deletion in frequency domain is "real" deletion.

Method

FreqX models neurons in the frequency domain, defining extract/filter operations based on vector norm changes and projections. It transforms neuron weights into a benchmark, calculates a "degree" for each neuron's effect on a sample, and aggregates these layer-by-layer.

In practice

• Use FreqX to analyze PFL model contributions.
• Apply frequency-domain deletion for faithful explanations.
• Adjust augmentation coefficient ε for explicit visualizations.

Topics

• Personalized Federated Learning
• Model Interpretability
• Frequency-Domain Analysis
• Feature Attribution
• Concept-Based Explanations
• Signal Processing

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• “Features” aren’t always the true computational primitives of a model, but that might be fine anyways — Features in mechanistic interpretability exist on a spectrum from memorization to computational primitives.
• Influcoder: Distilling Decoders' Gradient Influence Rankings into an Encoder for Data Attribution — Influcoder enables scalable, cost-effective data attribution for LLMs by distilling decoder influence into an encoder.
• xFODE: An Explainable Fuzzy Additive ODE Framework for System Identification — xFODE provides interpretable system identification by integrating fuzzy additive models and incremental state definitions with deep learning.
• Identifying Interactions at Scale for LLMs — SPEX and ProxySPEX efficiently identify influential interactions in ML models by exploiting sparsity and hierarchy.
• Towards Rigorous Explainability by Feature Attribution — Symbolic XAI methods offer rigorous alternatives to non-symbolic approaches for reliable feature importance attribution.
• Beyond Accuracy: Introducing a Symbolic-Mechanistic Approach to Interpretable Evaluation — Mechanism-aware evaluation using symbolic rules and interpretability better distinguishes true generalization from pattern exploitation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.