- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
The paper addresses proactive protection of publicly released LoRA adapters against unauthorised merging, and makes two claims.
First, that proactive defences designed for full-parameter release (PaRaMS, MergeLock, MergeBarrier) do not transfer to the fixed-base LoRA setting. The argument is structural: inside a LoRA factorisation the natural port A'=TA, B'=BT^{-1} leaves the induced update dW=(alpha/r)BA exactly unchanged (Eq. 9), so it is invisible to any merging method that consumes dW; applying the paired maps directly to dW does change the merge but breaks the standalone model, because the frozen base prevents cancellation (Eq. 26). Appendix A adds that MergeBarrier's attention and FFN components are not instantiable at the released rank (Eqs. 20, 22), and that realising a full-weight transformation with a rank-r update would require a base correction of rank at most 2r, while the correction actually induced can have rank d > 2r (A.3).
Second, it proposes LoRA-PBH, a release-time defence posed as a constrained max-min problem (Eq. 12): maximise the worst-case preservation barrier -- the increase in defender-task loss from the protected standalone model to the merged model -- subject to retaining standalone performance. Since the attacker's adapter and merge configuration are unknown, the perturbation set is approximated by probes built from defender-generated surrogate adapters (k-means pseudo-tasks warm-started from the defender adapter), the inner minimum is replaced by a differentiable SoftMin, and drift is limited by a normalised Frobenius anchor. Proposition 1 gives a conditional lower bound under a finite coverage radius and local Lipschitz assumption.
Evaluation covers ViT-B/32, B/16 and L/14 (five datasets, 20 defender-free-rider combinations each), three 7-8B LLM backbones with third-party released adapters, and qualitative SDXL results. On ViT-B/32 under task arithmetic at c=0.5, defender-task merged accuracy falls from 74.0% to 3.0% with standalone reported as +0.5%; protection transfers to six further merging methods, across c in [0.3, 0.8] and for two to five adapters. On LLMs merged accuracy falls 76.7% to 60.2%. A white-box adaptive attack recovers 3.0% to 34.4%. A runnable ViT-B/32 codebase is included.
Critical strengths
-
The structural diagnosis is correct and the appendix does the harder half of the work. The core identity is elementary, but Appendix A goes further: MergeBarrier's two components are not merely ineffective but not representable at the released rank (Eqs. 20, 22), and A.3 gives a counterexample where the required base correction has rank d > 2r. Figure 3 confirms both routes empirically for all three defences under a controlled fixed-rank budget. This is a checkable negative result and the part of the paper I would most want the community to have.
-
The protocol is broad and disciplined, and the released code backs it up. Each defender is protected once using defender-side data only, then evaluated against four unseen free-riders; six further merging methods are applied with no method-specific retraining; the coefficient sweep, the two-to-five adapter sweep, the alternative-probe study and the update-shift diagnostic each probe a distinct failure mode. Static inspection of src/method.py indicates agreement with Algorithm 1: the detached TA probe, the two-pass stop-gradient SoftMin weighting, the normalised anchor and the gradient surrogate are all implemented as described.
Critical weaknesses
-
A formally published, closely related prior work is missing, and with it a key LoRA-native baseline. TRAP2 (Jang et al., "Making Models Unmergeable via Scaling-Sensitive Loss Landscape", ICML 2026, PMLR 306:50690-50717; arXiv:2601.21898, public since January 2026) states the same base-cancellation obstacle that Section 5 analyses -- "applying paired transformations solely to the adapter dW fails to cancel terms associated with the frozen base model W0" -- already uses LoRA variants of PaRaMS and Merge-Lock as baselines, and proposes a training-time adapter-native objective J = L_nominal - lambda * L_off that shapes the loss against off-nominal rescalings. Its two adaptations are not identical to this paper's two routes, and I am not claiming the contributions coincide: the formal rank analysis, the seven-method coverage and the adaptive attack here are additional. But the submission currently compares only against its own adaptations of full-model defences, which it then shows to fail, so no LoRA-native defence appears in the comparison at all. Positioning and experiments both need revisiting.
-
Two results are framed in a way that omits the costs they carry. (a) Standalone preservation is discussed against only one of two valid baselines. The no-barrier ablation in Figure 5 -- same 100 steps, same data, Ls + Aw only -- reaches 86.4%, against 81.6% for the original adapter and 82.1% for full LoRA-PBH. Relative to the released adapter the method is +0.5, so the claim that it preserves existing utility stands; relative to simply continuing to optimise it is -4.3, which is the utility-versus-protection opportunity cost. Both numbers are in Figure 5; what is missing is any discussion of the difference. (b) The protection is blunter than the formulation suggests. Protected merged accuracies in Table 1 (0.5, 2.1, 9.3, 0.7, 2.5) sit near or below the uniform-chance reference implied by the paper's own class counts (0.51, 2.13, 10.00, 2.33, 2.22; RESISC45 the one case slightly above), and Figure 2 shows protected merged SDXL outputs as noise rather than as images that have lost the defender's style. For a defence meant to deny unauthorised capability retention, reaching chance is not itself a failure. The issue is scope: Table 10 shows the free-rider task also collapsing (74.0% to 11.7%) under TA at c = 0.5, so in the configuration the paper leads with, merging destroys both capabilities. Whether an owner can block unauthorised merging while still permitting sanctioned composition is left unanswered, and that gap belongs in the paper as a scope limitation.
Questions for authors
Items 1-3 would move my score; the rest are secondary.
-
(Experiment request) Please position the paper against TRAP2 (ICML 2026, PMLR 306:50690-50717; arXiv:2601.21898) and compare on the ViT-B/32 protocol under matched base model, data and standalone-utility constraint. The candidate increment over a pure scaling proxy is the surrogate-task probe, so please make it measurable: LoRA-PBH with its surrogate probe, LoRA-PBH with the surrogate replaced by scalar rescaling of dW_def, and TRAP2, with hardening cost. Please also report the relative norm and cosine similarity between dW_psi and dW_def, since warm-starting does not by itself bound how far the surrogate moves. TRAP2 Section 6.4 evaluates post-merge SFT recovery; how does LoRA-PBH fare there?
-
(Clarification) Please discuss both standalone baselines rather than one: +0.5 against the original adapter, and -4.3 against the no-barrier ablation of the same length. How should an owner weigh the opportunity cost when deciding whether to harden?
-
(Experiment request) The coefficient sweep uses a single SHARED coefficient, but task arithmetic does not require this. What happens under asymmetric coefficients, say lambda_def = 1.0 with lambda_atk in {0.1, 0.3, 0.5}? Separately, every task used as private defender data is a public dataset, so a free-rider could fine-tune the merged model on a disjoint split of it. Does the barrier survive either attack, on ViT and on LLMs?
-
(Clarification) Which configuration produced the reported tables? The released code differs from the paper in three places: Appendix D.1 says 30 epochs per vision adapter while config.py sets 35 / 56 / 22 / 21 / 25; the hardening coefficient is sampled from [0.3, 0.8] in code versus [0.2, 1.0] in the paper; and n_candidates=8 with keep=8 means no barrier filtering, versus R=16 filtered to J=8.
-
(Clarification) Two framing points. (a) With M=1 the SoftMin reduces to a single barrier and beta has no effect, so the objective optimised is a stochastic estimate of the expected barrier rather than the worst case of Eq. (12). Proposition 1 may well hold on a fixed model and fixed probe set; what it does not do is establish a worst-case guarantee for the resampled-per-step procedure actually run. Would you add an M>1 ablation or restate Section 6.3? Do you have any empirical handle on the coverage radius when the attacker's adapter is unconstrained? (b) On LLMs the protected merged accuracy (60.2%) is below the base model's own (69.2%) on every backbone, which arguably strengthens your case but makes the base model the right reference point; could Table 3 be reframed that way?
-
(Suggestion, optional) All runs use seed 0 with no variance reported, which matters given that protected standalone accuracy is reported as rising.
6 comments
We further analyze the surrogate updates. Averaged over runs and retained surrogates, their relative norm is , with a cosine similarity of 0.43 to the defender update. In comparison, actual free-rider updates have cosine similarities of only 0.02–0.07. Thus, the surrogate updates capture directions beyond simple rescaling of the defender update. The controlled ablation confirms that these additional directions are important for achieving the protection–utility trade-off.
(3) Post-merge fine-tuning.
Following the post-merge fine-tuning evaluation of TRAP2, we evaluate whether fine-tuning can recover the protected capability. We fine-tune merged models (TA with scaling ) for 200 AdamW steps (learning rate ), using either 64 labeled samples per task or a larger split containing 20% of each test set. We evaluate on the held-out remainder. Results are averaged over the 20 defender–free-rider pairs.
| Model | Before FT | 64 samples per task | 20% split |
|---|---|---|---|
| Unprotected merge | 74.0 | 67.8 | 76.2 |
| LoRA-PBH merge | 3.0 | 53.7 | 62.1 |
| Free-rider adapter only | 56.4 | 63.8 | 75.1 |
| Base model without adapter | 48.7 | 60.8 | 71.6 |
Although fine-tuning lifts the protected merge out of the initial performance collapse, it does not recover an advantage from retaining our adapter. The fine-tuned protected merge remains 10.1 and 13.0 percentage points below the free-rider adapter alone under the two data settings, respectively. It is also 7.1 and 9.5 points below the fine-tuned base model. In contrast, the unprotected merge performs better than both alternatives after fine-tuning.
These results suggest that under the evaluated fine-tuning budgets an attacker would obtain better defender-task performance by fine-tuning its own adapter without our protected adapter, or even the base model alone.
(4) Hardening cost.
All timings are measured using two RTX A6000 GPUs. LoRA-PBH starts from an existing trained adapter and takes 4.1 minutes in total for 100 hardening steps, 6.0 minutes for 200 steps, and 9.3 minutes for 400 steps. For reference, standard training of our adapter takes 45.2 minutes. TRAP2 is applied during adapter training. Under its official early-stopping protocol, one run takes 148.3 minutes (24k steps) for and 381.5 minutes (54k steps) for . Selecting requires one training run per candidate. Although the workflows are different, these results demonstrate the practical advantage of applying LoRA-PBH to an existing adapter without repeating its original training.
Q2: Standalone utility and opportunity cost
We thank the reviewer for pointing out the importance of considering both standalone baselines. We agree that the no-barrier ablation provides an important reference for understanding the opportunity cost of protection.
The two baselines answer different questions. The original adapter represents the owner's actual alternative, which motivates our standalone preservation constraint. The no-barrier control measures what the same optimization budget would achieve if spent exclusively on standalone utility. We therefore added a paired hardening-budget study, in which each LoRA-PBH run is compared with a no-barrier control using the same number of steps and random seed. Results are averaged over five runs.
| Hardening steps | No-barrier control | LoRA-PBH | vs. original | vs. control (paired) | Merged Defender Acc. | Total time |
|---|---|---|---|---|---|---|
| 100 | 86.5 | 82.2 | 3.1 | 4.1 min | ||
| 200 | 86.7 | 82.7 | 3.7 | 6.0 min | ||
| 400 | 88.1 | 84.2 | 3.5 | 9.3 min |
Protection remains at chance level at every budget (3.1–3.7%), while the standalone accuracy of LoRA-PBH remains above that of the original adapter. Compared with continued standalone optimization, the opportunity cost stays at approximately four percentage points ( to ), in exchange for reducing merged defender-task accuracy from 74.0% to 3.1–3.7%.
We believe this provides a practical basis for an adapter owner to evaluate whether hardening is worthwhile. For an already-trained or third-party adapter, LoRA-PBH preserves the original utility, with only a few additional minutes of hardening. In our LLM experiments, which use third-party released adapters, hardening changes standalone accuracy by only to points (Table 3). For an owner who can continue training, we recommend first training the adapter to convergence and then applying LoRA-PBH. The revised manuscript now discusses both baselines and this opportunity cost (Section 7.2.1 and Appendix D.5).
Q3: Asymmetric merging coefficients and fine-tuning with disjoint defender data
We thank the reviewer for suggesting these two important attack scenarios. The original coefficient sweep uses a shared merging coefficient, whereas an attacker can independently control the contribution of each adapter. Furthermore, since the defender datasets are public, an attacker may obtain additional samples that were not used during protection. We therefore extend our evaluation to asymmetric merging and fine-tuning with disjoint public data.
(1) Asymmetric merging coefficients.
We fix the defender coefficient to and vary the attacker coefficient over $\lambda_{\mathrm{fr}}\in\lbrace 0.1,0.3,0.5\rbrace $. The following table reports defender-task / free-rider-task accuracy after TA merging on ViT-B/32, averaged over 20 pairs and five runs.
| Unprotected | LoRA-PBH | |
|---|---|---|
| (0.5, 0.5) | 74.0 / 74.0 | 3.1 / 12.0 |
| (1.0, 0.1) | 80.2 / 53.9 | 62.7 / 13.8 |
| (1.0, 0.3) | 79.8 / 65.5 | 50.1 / 25.2 |
| (1.0, 0.5) | 74.6 / 68.3 | 28.6 / 19.4 |
Asymmetric coefficients do not give the free-rider a usable merge. With our adapter kept at full weight, the merged model moves closer to the defender's standalone configuration, so defender-task accuracy is higher than under equal coefficients. Nevertheless, compared with the unprotected merge under the same coefficients, LoRA-PBH reduces defender-task accuracy by 17.5–46.0 points and free-rider-task accuracy by 40.1–48.9 points.
More specifically, for the free-rider task, accuracy remains at 13.8–25.2% for all three attacker coefficients, substantially below the base model without any adapter (48.7%). For the defender task, accuracy decreases from 62.7% to 28.6% as the attacker coefficient increases. At , it is already close to the base-model level (50.1% vs. 48.7%), and at , it is 20.1 points below the base model.
We further evaluate the same asymmetric configurations on three LLM backbones. The following results are averaged over backbones and pairs. Each entry reports defender-task / free-rider-task accuracy.
| Unprotected | LoRA-PBH | |
|---|---|---|
| (0.5, 0.5) | 76.7 / 76.7 | 60.2 / 68.8 |
| (1.0, 0.1) | 81.3 / 70.8 | 68.2 / 65.9 |
| (1.0, 0.3) | 80.2 / 73.2 | 66.3 / 67.2 |
| (1.0, 0.5) | 78.6 / 75.4 | 63.5 / 68.6 |
With , both tasks remain below the average base-model accuracy (69.2%) for all three attacker coefficients. Defender-task accuracy is 63.5–68.2%, compared with 82.1% for the protected adapter used alone. Free-rider-task accuracy is 65.9–68.6%, compared with 70.8–75.4% without protection. Thus, the asymmetric attack does not produce a merge that preserves both capabilities on either ViT or LLM backbones. The revised Section 7.3 states this result, and Appendix D.4 reports both tables.
(2) Fine-tuning with disjoint public data.
We also evaluate an attacker who obtains additional labeled samples from the same public defender dataset but does not have access to the samples used during LoRA-PBH hardening. Such an attacker holds labeled defender-domain data: it could train its own adapter and has no need to free-ride, which is why the threat model in Section 4.3 excludes it. We nevertheless measured this case as requested. We use disjoint data splits for protection and attacker fine-tuning, and evaluate on held-out data.
For ViT-B/32, the fine-tuning results are reported in the table under Q1. Although fine-tuning recovers part of the defender capability, the protected merge remains inferior to fine-tuning the free-rider adapter alone or the base model without our adapter.
We further conduct the disjoint-data fine-tuning experiment on the three LLM backbones. The same trend is observed: relative to the free-rider adapter alone, fine-tuning a merge that retains our protected adapter incurs an additional performance cost of 3.1 points with 64 samples and 1.9 points with the 20% split. In contrast, including the unprotected adapter provides gains of 3.7 and 5.2 points under the respective settings.
These results indicate that, even when the attacker obtains additional public defender-domain data that does not overlap with the protection data, retaining the protected adapter does not provide a measurable advantage over fine-tuning without it. The revised Section 4.3 states the attacker's resources explicitly, and Appendix D.3 reports these results.
We sincerely thank the reviewer for the detailed and constructive review. We particularly appreciate the suggestions for strengthening our empirical evaluation and clarifying the scope of our protection mechanism. Following these suggestions, we have conducted additional experiments and revised our presentation. We first address the three questions that the reviewer identified as most important for the evaluation, followed by additional clarifications.
Q1: Positioning against TRAP2 and the requested comparison
We thank the reviewer for pointing us to TRAP2. The revised manuscript cites and discusses TRAP2 in Sections 1 and 2, and adds the comparison below to Section 7.2.1 and Appendix D.7. The primary advantage of LoRA-PBH is its post-hoc applicability: it can protect an already-trained LoRA adapter without modifying or repeating its original training process, whereas TRAP2 incorporates its protection objective during adapter training. This makes LoRA-PBH applicable to existing adapters, including third-party adapters that have already been released.
Beyond this difference in applicability, our protection objectives also differ. TRAP2 makes an adapter sensitive to rescaling its own update, using scaling variations as a proxy for merging. In contrast, LoRA-PBH explicitly models the perturbation introduced by merging with an external adapter. Since the actual merging partner is unknown, we construct defender-generated surrogate adapters to approximate such external updates and directly optimize the preservation barrier under these perturbations. This allows LoRA-PBH to model effects beyond simple rescaling of the defender's own update while preserving standalone utility. LoRA-PBH is therefore not a modification of TRAP2: both methods shape the loss landscape around the released update, but they model different perturbations and act at different stages. TRAP2 notes that paired transformations on the adapter cannot cancel the frozen base; Section 5 and Appendix A formalize this obstacle (Eq. 26 and the rank result of Appendix A.3) and additionally cover the factor-internal route, which leaves ΔW unchanged (Eq. 9), as the revised Appendix D.7 now states.
(1) Direct comparison with TRAP2.
We implemented TRAP2 using its official code and hyperparameters (LoRA on attention projections, r=16, with early stopping), evaluating both values in its hyperparameter grid, $\lambda\in\lbrace 0.001,0.01\rbrace $. All adapters start from the same CLIP ViT-B/32 weights and are evaluated using the same pipeline, over the 20 defender–free-rider pairs of Table 1, with Task Arithmetic (TA) at c=0.5. The scaling-probe variants replace the surrogate update with a rescaling of the defender update, using the same 100-step hardening budget. Results are averaged over five random seeds, so the 100-step LoRA-PBH row differs slightly from the seed-0 values of Table 1 (82.1, +0.5, 3.0).
Both TRAP2 and LoRA-PBH reduce merged defender-task accuracy to approximately chance level (3.1–4.8%). However, their standalone utility differs substantially. The TRAP2 adapters end 5.9–10.0 percentage points below our unprotected adapters, whereas LoRA-PBH preserves the standalone performance of the adapters it hardens (+0.6 points with 100 steps and +2.6 points with 400 steps). On the free-rider task, the merged model retains 23.2–25.7% with TRAP2 and 12.0–12.3% with LoRA-PBH; we discuss this effect under W2(b).
(2) Why surrogate probes matter.
To isolate the contribution of surrogate probes, we replace them with two scaling-based alternatives while keeping the original adapters and 100-step hardening budget unchanged.
Neither scaling-based alternative achieves both protection and standalone utility. When ΔWψ=ΔWdef, standalone accuracy remains relatively close to the original (78.3%), but merged defender-task accuracy remains high (69.7%). Conversely, when ΔWψ=0, merged accuracy drops to 3.8%, but standalone accuracy collapses to 53.2%. In contrast, the surrogate-based objective achieves both objectives simultaneously, retaining 82.2% standalone accuracy while reducing merged accuracy to 3.1%.