- Rating
- 4
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes Gradient Sealing Unlearning (GSU), a preemptive defense against the reacquisition of forbidden capabilities through downstream fine-tuning. The method consists of three stages: exposing potential acquisition pathways through proxy fine-tuning, localizing sensitive activation channels, and sealing these channels by pushing their pre-activations toward the negative saturation region of the SiLU activation function. The idea is interesting, and the experimental evaluation covers multiple model architectures, benchmarks, and existing defense methods.
Critical strengths
- The paper distinguishes a model’s current capabilities from its ability to acquire them through fine-tuning, offering a valuable perspective on LLM unlearning.
- The paper provides several complementary analyses supporting the proposed gradient-sealing mechanism.
Critical weaknesses
- The method uses a soft penalty to push selected pre-activations toward a region with small derivatives. However, “sealing” could be mistaken for freezing weights or removing neurons. Please clarify that it changes the responses of selected gates without imposing hard constraints on parameter updates.
- The evaluation clearly specifies that released models must retain at least 95% of the reference utility, while usable attack checkpoints must retain at least 90%. A brief explanation of why a slightly lower threshold is allowed after fine-tuning would make the protocol easier to interpret. This is a request to clarify the existing protocol choice.
Questions for authors
- Could the authors clarify how gradient sealing differs from freezing or removing neurons? Specifically, does sealing use a soft loss penalty to modify gate responses while keeping all model parameters trainable?
- What motivates the utility thresholds of 95% for released models and 90% for usable attack checkpoints? Is the lower attack-time threshold intended to accommodate some utility loss during fine-tuning while excluding attacks that severely degrade model utility?
2 comments
Please engage in the discussion with authors
Dear Reviewer Nx7q,
Please engage in the discussion with authors, thank you!
AC
Response to Reviewer Nx7q
Thank you for your positive review, clear acceptance recommendation, and recognition of our problem framing, simple design, and broad evaluation. We address both clarification points below and will include them in the revision.
Weakness 1: What “sealing” changes: soft gate-response control, not freezing or removal
We will clarify in §4.3 that sealing is a soft, training-time penalty on gate responses—not freezing, masking, or pruning:
Soft hinge penalty during defense only: Lseal (Eq. 4) applies max(uℓji(θ)−τ,0)2 to the sparse localized gate set G (1%–10% of channels in K∈{1,2,4} layers), nudging forbidden-input pre-activations into the SiLU negative tail (u≤τ<0) with zero direct gradient once u≤τ and no hard clamp or weight mask.
All weights stay 100% trainable during attack: Full-parameter AdamW updates every parameter freely with no sealing loss or frozen units. Across 30 epochs (390 updates) on Qwen3.5-2B:
Table C.1: Gate sensitivity and parameter updates during 30-epoch attack.
Selected-channel weights actively move at every step (slightly larger RMS than control), while their mean |SiLU'(u)| remains 11.47% lower (0.391546 vs. 0.442273).
No neurons removed; utility stays intact: GSU preserves 98.05% of reference utility after 30 attack epochs (100.18% at release in Table A.1) and maintains benign fine-tunability (Fig. 4c).
Weakness 2: Why released models use 95% utility and attack checkpoints 90%
The two thresholds govern different roles in the attack–defense game (Defs. 2.1–2.2):
Question 1: How does gradient sealing differ from freezing or removing neurons?
Yes (see W1, Table C.1): sealing uses a soft hinge penalty only during defense training, becoming inactive once u≤τ with no hard constraints or pruning. During downstream attacks, full-parameter AdamW updates every weight normally (RMS 6.077–8.249×10⁻⁸), while targeted gates retain an 11.47% lower mean |SiLU'(u)| (0.391546 vs. 0.442273) and 98.05% reference utility.
Question 2: What motivates the 95% and 90% utility thresholds?
Exactly (see W2): 95% enforces high utility on the defender's release, while 90% filters attacker checkpoints in UA90—accommodating minor fine-tuning drift while excluding destructive attacks. Since our 30-epoch checkpoints retain >98% utility, all reported UA90 results sit well above both bars.