New Fingerprinting Framework ‘ForgetMark’ Enables Stealthy Model Ownership Verification

Global: New Fingerprinting Framework ‘ForgetMark’ Enables Stealthy Model Ownership Verification

A team of AI researchers announced in January 2026 that they have developed a novel model‑fingerprinting technique called ForgetMark. The approach, detailed in a paper posted on the arXiv preprint server, aims to provide a reliable way to prove ownership of neural network models while addressing limitations of prior invasive backdoor methods.

Background and Motivation

Existing fingerprinting methods often rely on high‑perplexity triggers that can be easily filtered by heuristic detectors, and they may produce fixed response patterns that expose the fingerprint. ForgetMark was designed to overcome these challenges by encoding provenance through targeted unlearning rather than static triggers.

Methodology

The framework constructs a compact, human‑readable key‑value set using an auxiliary assistant model and a predictive‑entropy ranking scheme. Lightweight LoRA adapters are then trained to suppress the original values associated with each key while preserving the model’s overall capabilities.

Verification Process

Ownership verification can be performed under black‑box or gray‑box access. The method aggregates likelihood scores and semantic evidence to compute a fingerprint success rate, allowing verifiers to assess ownership without requiring full model internals.

Stealth and Robustness

Because ForgetMark relies on probabilistic forgetting traces instead of fixed trigger‑response patterns, it reduces detectability and minimizes false activations on benign inputs. The authors report that the technique is less susceptible to detection by existing heuristic detectors.

Empirical Results

Across a range of architectures and experimental settings, the authors observed 100% ownership verification on fingerprinted models while maintaining standard performance metrics. The approach also outperformed traditional backdoor baselines in terms of stealthiness, demonstrated robustness to model merging, and remained effective after moderate incremental fine‑tuning.

Open‑Source Release

The research team has made the code and associated data publicly available on GitHub, inviting further evaluation and adoption by the broader community.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.