TraceMark: Chain-Consistent Parameter Watermarking for Lifecycle-Aware Ownership Verification
Year: 2026
Type: article
Abstract: Deep visual models are valuable intellectual-property assets, but existing watermarking methods usually verify ownership on a single static checkpoint. This setting is insufficient for deployment pipelines in which models are repeatedly updated, fine-tuned, pruned, quantized, and redistributed. We propose TRACEMARK, a chain-consistent parameter water-marking framework for lifecycle-aware ownership verification. TRACEMARK conditions each stage code on a deterministic projection of the predecessor checkpoint, so the embedded evidence records structured lineage rather than an isolated terminal mark. It further uses keyed importance-aware carrier selection, progressive stage-wise embedding, and a two-level verification protocol that separates watermark existence from predecessor consistency. Experiments on CIFAR-10, CIFAR-100, Tiny ImageNet, and GTSRB across multiple CNN and Vision Transformer architectures show that TRACEMARK preserves main-task performance with an average accuracy drop of 0.43 percentage points, maintains strong recoverability under fine-tuning, unstructured pruning, and weight-only quantization, and rejects incorrect predecessor claims with substantially lower false acceptance than static or non-chain baselines.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,645
Awards: G220962924, G4840164131, G6112228076
OpenAlex ID: W7204226403
