Answers are grounded in this paper and cite their sources.
Framework detects model poisoning in machine learning systems using forensics, indicating scalable defenses.
As machine learning (ML) models are increasingly deployed in high‐stakes and security‐sensitive environments, the risk of model poisoning—where adversaries imperceptibly manipulate training data to subvert model behavior—poses a growing challenge. In this work, we present a unified forensic analysis framework designed to detect such threats by examining internal model behavior rather than relying solely on input anomalies. Our novel pipeline combines structural, geometric, statistical, and interpretability‐driven methods—including model inversion, topological data analysis, Shapley value attribution, and Benford's Law analysis—to reveal latent signals of adversarial manipulation. Distinctively, we extend the evaluation beyond traditional convolutional neural network architectures by applying our forensic analysis framework to ResNet‐18 and YOLOv8, demonstrating its effectiveness across varied ML model families. Using comprehensive benchmarks on the CIFAR‐10 and CelebA datasets, we show that our framework achieves high detection accuracy while offering interpretable forensic insights into model behavior. This approach provides a scalable, robust defense mechanism for securing ML models against poisoning attacks in mission‐critical applications.
No takes yet. Share an insight, caveat, or question.
Monkam et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: