Frequency-Aware Model Parameter Explorer: A new attribution method for improving explainability

Researchers introduce FAMPE, a novel attribution method that uses frequency-domain analysis to improve explainability in deep neural networks. By separately perturbing high and low-frequency components through FFT-based techniques, the method outperforms existing attribution approaches on ImageNet across multiple architectures without requiring manual baseline selection.

FAMPE addresses a fundamental limitation in current attribution methods: their inability to capture fine-grained frequency information when generating adversarial samples. Traditional approaches apply uniform perturbations across the entire frequency spectrum, losing critical details that neural networks actually rely on for accurate predictions. This research demonstrates that selective frequency manipulation provides a more precise window into model decision-making by directly probing which spectral features influence outputs most.

The technical innovation centers on integrating frequency-aware exploration into parameter-level attribution—a connection unexplored in prior work. By using weighted perturbations that separately modulate high and low-frequency components through an energy-driven spectral cutoff, FAMPE generates more informative attribution maps without arbitrary baseline selection, a common pain point in explainability research. The method's empirical validation across ImageNet using multiple architectures (CNNs and Vision Transformers) provides robust evidence of its effectiveness.

These findings carry implications for AI transparency and trustworthiness. As neural networks increasingly influence critical decisions, understanding which features drive predictions becomes essential for safety, debugging, and regulatory compliance. The discovery that low-frequency-dominated images systematically benefit from high-frequency perturbations suggests adaptive spectral exploration could further improve attribution precision. The 4-12% performance improvements over existing methods, combined with ablation studies confirming high-frequency perturbations' importance, establish FAMPE as a meaningful advancement in model interpretability research.

• →FAMPE uses FFT-based frequency-aware perturbations to improve neural network attribution and explainability
• →Method outperforms existing approaches by 4-12% on major architectures without requiring manual baseline selection
• →High-frequency perturbations prove disproportionately important for attribution precision versus low-frequency components
• →Adaptive spectral exploration shows promise for further improvements on frequency-specific image categories
• →Integration of frequency analysis into parameter exploration establishes a novel connection in explainability research