WFR-Guard: detecting poisoned training patterns via watermark anchors and randomized feature projections

Pattern recognition systems are now being deployed in high-stakes environments where data integrity is critical, but often untrusted. Data-poisoning attacks are a serious threat, allowing attackers to inject malicious samples that change the behavior of the model while keeping the validation accuracy seemingly unchanged. We propose WFR-Guard, a security-motivated, model-agnostic defense framework that combines cryptographic watermarking (WM) with randomized feature verification. WFR-Guard embeds imperceptible least significant bit (LSB) watermarks into trusted data subsets to enable verifiable provenance, while epoch-wise random projection disrupts adversarial alignment without requiring changes to the classifier architecture or a separate post-hoc retraining stage. We formalize a cryptographic security game, and bound the adversary’s success probability to below $$3\%$$ . Extensive experiments on CIFAR-10, GTSRB, and PathMNIST under label flipping (LF), feature collision (FC), BadNets backdoor (BD), blended trigger (BT), and adaptive mimicry (AM) attacks demonstrate that WFR-Guard maintains $$\ge 89\%$$ detection accuracy (DA) even at $$50\%$$ poisoning intensity, while preserving clean-data accuracy up to $$98.5\%$$ . The defense achieves low FPR, strong white-box robustness, and lightweight inference overhead ( $$\sim 2.3$$ ms per image). These results establish WFR-Guard as a practical, verifiable solution for securing high-confidence pattern recognition systems against a wide range of data-poisoning attacks. The framework is particularly relevant for safety-critical applications where corrupted training data can lead to severe operational consequences, including intelligent transportation systems and AI-assisted healthcare platforms.