RESEARCH · RESEARCH · #3185
Embedding-space fingerprinting recovers fine-tune domain from LoRA artifacts
This paper (arXiv:2611.07610v1) studies whether the domain of a LoRA fine-tune can be recovered from the model artifact alone. Using Qwen3.5-0.8B as the base, the authors produce 36 LoRA fine-tunes across 18 public datasets in five domains and extract three feature families (weight deltas, internal activations, and outputs) under a pre-registered leave-one-dataset-out (LODO) protocol; a classifier identifies the domain well above chance from gray-box activations (61%) and black-box outputs (64%), while white-box weight statistics reach 42%, and the black-box result holds even when only generated text is used.
KEY POINTS
- This paper (arXiv:2611.07610v1) studies whether the domain of a LoRA fine-tune can be recovered from the model artifact alone.
- Using Qwen3.5-0.8B as the base, the authors produce 36 LoRA fine-tunes across 18 public datasets in five domains and extract three feature families (weight deltas, internal activations, and outputs) under a pre-registered leave-one-dataset-out (LODO) protocol; a classifier identifies the domain well above chance from gray-box activations (61%) and black-box outputs (64%), while white-box weight statistics reach 42%, and the black-box result holds even when only generated text is used.
- Shows that fine-tuned LoRA artifacts can leak the training-data domain from weights, activations, or outputs, with implications for privacy, model sharing, and defenses such as watermarking.
WHY IT MATTERS
Shows that fine-tuned LoRA artifacts can leak the training-data domain from weights, activations, or outputs, with implications for privacy, model sharing, and defenses such as watermarking.