acceptodds
Under review as a conference paper at ICLR 2027

Differentially Private Multimodal In-Context Learning

Abstract

Vision-language models (VLMs) are increasingly deployed on sensitive data such as medical scans and financial documents, where in-context learning offers customization without retraining. Yet existing differentially private in-context learning methods are limited to few-shot, text-only settings because privacy cost scales with the number of demonstrations, and a single image can consume hundreds of tokens. We introduce Differentially Private Multimodal Task Vectors (DP-MTV), the first framework for many-shot multimodal in-context learning with differential privacy. Rather than protecting individual demonstrations, DP-MTV aggregates hundreds of private image-text examples into a single steering vector and privatizes the aggregate. Privacy cost is paid once at construction; the released vector then answers unlimited queries at zero additional cost. Across eight benchmarks and three VLM architectures, DP-MTV matches or exceeds non-private performance on the majority of tasks at epsilon=2, and on classification consistently outperforms the non-private baseline. Beyond this public-data setting, a fully-private variant that uses no auxiliary data for head selection matches it at epsilon=5, (for example, 96% on Flowers-102 with Qwen-VL versus 77% for non-private MTV)with a per-layer selection scheme extending competitive accuracy to tighter budgets. A novel membership inference attack validates these guarantees: 100% success without protection drops to the 50% random-guess baseline under DP-MTV.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
differential privacy, in-context learning, vision-language models, task vectors, membership inference attacks, activation steering, multimodal learning, privacy-preserving machine learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_CjQUOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Differentially Private Multimodal Task Vectors (DP-MTV), a framework for enabling privacy-preserving many-shot multimodal in-context learning. Building on Multimodal Task Vectors (MTV), the method aggregates attention-head activations from private image-text demonstrations into a compact representation and applies calibrated Gaussian noise to provide differential privacy. The authors develop two variants: one using public auxiliary data for attention-head selection and another performing head selection privately through sample-and-aggregate and the exponential mechanism. The paper also introduces an activation-similarity membership inference attack to evaluate privacy leakage. Experiments across eight benchmarks and three vision-language models demonstrate promising privacy–utility tradeoffs, particularly for the public-data variant and fine-grained classification tasks. The framework enables repeated inference using the privatized task vector without additional privacy expenditure.

Critical strengths
  1. The paper addresses an important challenge in multimodal in-context learning: leveraging private image-text demonstrations while protecting individual records. Aggregating demonstration information into a reusable task vector provides an appealing alternative to query-dependent privacy mechanisms. The separation between offline private construction and online inference is particularly valuable, as the privatized representation can be reused without additional privacy expenditure through the post-processing property of differential privacy.

  2. The paper evaluates DP-MTV on eight benchmarks spanning visual question answering and fine-grained classification, using three different VLM architectures. It investigates both public-data and fully-private head-selection variants, reports privacy–utility tradeoffs under different privacy budgets, and includes an empirical membership inference analysis. The strong classification results suggest that activation clipping and noisy aggregation may also provide useful regularization effects beyond privacy protection.

Critical weaknesses

1. Potentially incorrect sensitivity analysis and insufficiently specified privacy accounting.

The sensitivity analysis in Theorem 4.1 and Appendix B appears to overlook an important factor. The authors claim that the sensitivity of the aggregated task vector under record-level add/remove adjacency is bounded by ∣S∣C/m\sqrt{|S|}C/m. However, each chunk contains multiple demonstrations jointly processed by a nonlinear VLM. Adding or removing one demonstration may therefore change the entire activation representation of the affected chunk. Since both the original and modified chunk activations are independently clipped to norm CC per layer, the triangle inequality only guarantees that their difference is bounded by 2C2C, rather than CC. Consequently, even assuming that only one chunk is affected and the number of chunks remains fixed, a conservative sensitivity bound would be 2∣S∣C/m2\sqrt{|S|}C/m. The manuscript does not provide additional assumptions or arguments that justify removing this factor of two.

This concern is fundamental because the Gaussian noise scale is directly calibrated using the claimed sensitivity. If the sensitivity is underestimated, the reported differential privacy guarantees may not hold at the stated privacy budgets. The authors should clarify the precise neighboring-dataset construction and provide a rigorous justification for their sensitivity bound.

2. Insufficient experimental validation of the privacy mechanism and its claimed utility advantages.

The experimental results raise several concerns about whether the reported advantages can be attributed to the proposed privacy-preserving mechanism. In Table 2, DP-MTV substantially outperforms non-private MTV on several classification tasks, such as Flowers-102 on Qwen-VL (approximately 96% versus 77%). The authors attribute these improvements to activation clipping and Gaussian-noise regularization, but no controlled ablation separates the effects of clipping, noise injection, and head selection. In particular, a clipped but non-private MTV baseline is missing, making it difficult to determine whether the improvement comes from the DP mechanism or simply from modifying the original MTV implementation.

More importantly, the strong results primarily come from the public-data variant. Table 3 shows that the fully-private variant suffers severe performance degradation at tighter privacy budgets. At ϵ=2\epsilon=2, accuracy collapses to near zero on most tasks for ViLA and Idefics2, while the Qwen-VL results are also substantially degraded. This limitation considerably narrows the practical applicability of the proposed framework when suitable public auxiliary data is unavailable.

Finally, the privacy evaluation relies on a single activation-similarity membership inference attack. Algorithm 1 assumes access to task-vector representations and the ability to construct additional reference vectors using the underlying model, which differs from the output-access adversary described in the main threat model. Although the reported attack success decreases under DP-MTV, this result demonstrates resistance to one particular attack rather than independently validating the claimed formal privacy guarantee. The paper would benefit from a clearer threat-model definition, stronger attack evaluations, and controlled experiments that isolate the contributions of its individual components.

Questions for authors
  1. In Theorem 4.1, why is the sensitivity bounded by ∣S∣C/m\sqrt{|S|}C/m rather than 2∣S∣C/m2\sqrt{|S|}C/m? Since adding or removing a demonstration can change the entire chunk activation, how is this difference accounted for?

  2. Algorithm 2 requires disjoint chunks containing one target and KK demonstrations. With m=100m=100 and K=2K=2 for classification, why does Section 5 report only 100 private examples rather than 300? Are any private samples reused across chunks?

  3. If the sensitivity bound or chunk construction requires correction, how would the calibrated Gaussian noise and the reported privacy–utility results change?

  4. In Algorithm 3, the validation subset is sampled from the private dataset for exponential-mechanism selection. How does the privacy analysis account for changes in the validation subset under add/remove adjacency, particularly when the sampling procedure depends on dataset size?

  5. Can the authors compare DP-MTV against MTV with activation clipping but without Gaussian noise, and against MTV with noise but without clipping, to identify the source of the substantial classification gains?

  6. Table 3 shows severe accuracy degradation for the fully-private variant at ϵ≤2\epsilon \leq 2. What causes these failures, and can the proposed per-layer selection scheme consistently recover performance across all three model architectures?

  7. Algorithm 1 requires access to activation representations and the ability to construct reference task vectors. How does this assumption relate to the output-access adversary described in Section 4, and is the attack representative of realistic deployment settings?

  8. Can the authors evaluate additional membership inference strategies, including attacks based on model outputs or adaptive queries, rather than relying exclusively on the proposed activation-similarity attack?

4 comments

We thank the reviewer for recognizing that separating offline private construction from online inference lets the released vector be reused for unlimited queries at no further privacy cost. Point-by-point response below. This comment covers Concerns 1 and 2, and two further comments cover Concerns 3 to 6.

Concern 1: The sensitivity in Theorem 4.1 is a factor of two too small (W1, Q1, Q3).

The reviewer is right. The correct sensitivity is a factor of two larger than the bound in Theorem 4.1. Appendix B's bound holds only when the removed record is alone in its chunk. When the chunk holds other records, removing one changes the whole chunk's activation, and the old and new activations can point in opposite directions, so the change can reach twice the clipping bound. The sensitivity, which is the most one record can change the released vector and sets how much noise we add, is therefore Δ = 2√|S|·C/m, where |S| is the number of layers, C is the clipping size per layer, and m is the number of chunks. Our DP-SGD comparison in Appendix B does not apply, because DP-SGD clips each record separately, while a chunk combines several records.

A larger sensitivity means more noise at every privacy budget. We ran the public-data experiments on Qwen-VL with the noise set for the new bound, at the same budgets and with the same accounting as the submitted runs, so only the sensitivity changed. The trend is the same as in the paper, shifted to a slightly larger budget: DP-MTV still matches or beats non-private MTV on the same 5 of the 8 tasks (table below), and the same ordering of tasks holds, with the extra noise simply moving the useful range up. It falls short only on OK-VQA and TextVQA, where MTV itself gives little or no gain over zero-shot, as already seen in Table 2. At ε=1 the larger noise leaves too little signal on the VQA tasks, so we present ε=2 to 5 as the operating range.

Qwen-VL, public datazero-shotnon-private MTVDP-MTV ε=2DP-MTV ε=5
VizWiz355555.560.3
PathVQA252828.529.1
VQA-RAD38400.341.9
OK-VQA586043.046.4
TextVQA686447.344.9
Flowers547798.093.8
CUB577891.693.8
DTD547692.793.1

The zero-shot and non-private MTV columns are the Qwen-VL rows of Table 2.

For the camera-ready, Theorem 4.1, Algorithm 2 and Appendix B will use the new bound and state the full privacy accounting, including the random sampling of the private records it relies on. Tables 2 and 3 and Figure 3 will report all three models with the new noise, averaged over three seeds as in the paper.

Concern 2: The chunks in Algorithm 2 do not match the 100 classification examples in Section 5 (Q2).

The count of 100 is correct; the two tasks group records into chunks differently. In both tasks a chunk is a single forward pass whose last-token activation is clipped once. In classification, each chunk is one query image together with its two labeled support images, and the private record is the query image, so 100 chunks cover 100 records. In VQA, each chunk is one target plus eight demonstrations, so 100 chunks cover 900 records. The numbers looked inconsistent because Section 4 described a classification record as a bare image rather than as the query image inside its chunk.

The chunks are disjoint: each record belongs to exactly one chunk, and the number and size of the chunks are fixed before looking at the data, so adding or removing one record changes only its own chunk and leaves every other chunk unchanged. This is what limits the sensitivity to a single chunk, and it is also why that chunk's change can be two-sided: removing the record still leaves the rest of its chunk, which re-forms into a different clipped activation, giving the factor of two in Concern 1. We will make this record definition and chunk placement explicit in Section 4 and Theorem 4.1.

Continued from the previous comment. This comment covers Concerns 3 to 5.

Concern 3: The validation records in Algorithm 3 may break privacy when a record is added or removed (Q4).

Algorithm 3 has two private steps, and each stays private when a record is added or removed.

The first step releases θ, the average of each chunk's p head probabilities over c chunks. Each probability is between 0 and 1, so the biggest change one record can make is to move all p of them from 0 to 1. Its sensitivity √p/c already uses this worst case, so it covers any change to a chunk and needs no factor of two.

The second step picks the final mask with the exponential mechanism, scored on validation records drawn at random. These records are drawn without looking at their content. One record then changes at most one validation loss, and each loss is capped at C_sel, so one record moves a candidate's score by at most C_sel. This gives (ε_mask, 0)-DP.

The C_sel bound holds for the results we report. The validation set is drawn without looking at the private records, so adding or removing one record changes at most one score, whatever the dataset size. The reviewer is right that a cleaner way to state the same guarantee is to include each record independently with a fixed probability; this is only a restatement of the construction, not a change to the mechanism or to the numbers above, and we will adopt it in the camera-ready to make the C_sel bound self-evident.

Concern 4: The gain over non-private MTV may come from clipping or noise rather than the privacy mechanism (W2, Q5).

The gain over non-private MTV is on the classification tasks, which is where this ablation applies. To separate the three factors, we keep the selected heads fixed and vary only what is placed at them (Qwen-VL, public head selection, accuracy in %). MTV and DP-MTV use the same selected heads, so head selection is not what sets them apart; the difference is the vector placed there.

What is placed at the headsFlowersCUBDTD
nothing (zero-shot)545655
raw MTV vector797980
MTV vector, clipped, no noise959393
DP-MTV (clipped, plus noise)969393
ε=2 noise alone (no vector)959492
the same, but at random heads495050

The raw vector reaches 79 on Flowers, matching non-private MTV in Table 2 (77 and 78 there, averaged over three seeds), and all three tasks behave the same. The table reads off each factor in turn. Clipping the vector with no noise reaches DP-MTV's accuracy (95 vs 96), so clipping, not noise, produces the gain; adding the noise on top changes nothing (96 vs 95). The selected heads are also required: the same noise placed at random heads instead stays at chance. Noise and head selection are shared by MTV and DP-MTV, so the one step that differs, and that moves the number, is clipping.

Varying the clip size C alone, with no noise, makes this direct. As C grows and less is clipped, the clip-only accuracy falls smoothly back to raw MTV, which it reaches once C is large enough to clip almost nothing.

clip-only accuracyC=1C=8C=20C=100raw MTV
Flowers9593897979
CUB9393857979
DTD9393868080

Clipping is therefore the step that turns MTV into the stronger result, and relaxing it recovers plain MTV.

This gain is specific to the 2-way classification tasks, where non-private MTV is weak (77 to 78). On the VQA tasks, where MTV gives a real gain over zero-shot, DP-MTV matches rather than exceeds MTV, and the added noise has a real cost. For the camera-ready, this ablation will replace our explanation in Section 5, and we will add the matching VQA controls.

Continued from the previous comment. This comment covers Concern 5.

Concern 5: The fully-private variant collapses at small budgets (W2b, Q6).

Our main method uses public data to pick the heads at no privacy cost. The fully-private variant removes that assumption: it also chooses the heads from the private data, so it needs no public data at all. The price is that one budget now covers both steps, split three ways (ε_tv + ε_θ + ε_mask). At a tight budget each part is too small to find the useful heads, so accuracy collapses at ε=1.

The fix is to select whole layers instead of individual heads. This releases 32 numbers instead of 1024, so the private selection needs far less noise. At ε=2 it recovers accuracy across all three models (per-head vs per-layer, from the appendix granularity table).

Fully-private, ε=2VizWizOK-VQATextVQAFlowersCUBDTD
Qwen-VL, single heads112413454049
Qwen-VL, whole layers425665757383
idefics2, single heads231000
idefics2, whole layers173044697587
ViLA, single heads002000
ViLA, whole layers245495041

Per-layer selection beats per-head on every cell, often by wide margins. The comparison is about selection design, not the noise level: per-layer releases 32 values instead of 1024, so it wins at any budget. At ε=1 the noise is large enough to make the selection almost random on all three models, so neither choice helps there; the fix works at moderate budgets, not the tightest ones.

That leaves the main question: once the heads are chosen, is the variant actually useful? The table below gives the full Qwen-VL results across budgets. Accuracy rises with ε, which is the expected privacy-utility tradeoff when the whole method runs on private data under one budget: a tight ε leaves too little signal, and ε=5 is where it holds up.

Fully-private Qwen-VLzero-shotε=1ε=2ε=5
VizWiz43.39.731.749.5
PathVQA25.222.119.127.1
VQA-RAD40.50.00.041.5
OK-VQA57.517.639.254.4
TextVQA68.723.536.060.1
Flowers56.816.548.960.6
CUB58.552.153.676.9
DTD52.958.861.469.1

At ε=5 the fully-private variant reaches or passes zero-shot on 6 of the 8 tasks (all three classification tasks, VizWiz, VQA-RAD and PathVQA), with OK-VQA and TextVQA close behind. We therefore present it as reliable at ε=5, as an option for when no public data exists, with the public-data variant as the main method. For the camera-ready, Table 3, the granularity table and this table will use the corrected bound throughout, averaged over three seeds, with the two VQA tasks run again at the same chunk setting as the rest, and the abstract will describe the fully-private variant in these terms.

Continued from the previous comment. This comment covers Concern 6.

Concern 6: The attack uses activations, unlike the output-only threat model, and only one attack is tested (W2c, Q7, Q8).

Our attacker is stronger than one in deployment, not weaker. It sees the activations and uses the public model, so anyone holding the released vector and the public model can reproduce every output the steered model gives. The output-only attacker in Section 4 is therefore a weaker special case of ours, and defending against the stronger one implies defending against it. The paper's limitations already note this white-box assumption.

The attack itself is the standard membership test that differential privacy limits. The attacker is given two task vectors, both released by DP-MTV with independent noise: one built from demonstrations that include the target x* (τ_mem), one from the same demonstrations with x* replaced by another record (τ_non). It guesses that x* is in whichever vector is more similar to x*'s own activation a*. Exactly one vector contains x*, so a random guess is right half the time, and differential privacy drives the attacker's accuracy toward that 50% as ε shrinks, which is the trend in Figure 3.

We test this one attack because no existing membership attack fits this setting. Standard attacks read a trained model's confidence on its own training examples, but a private steering vector is not trained that way, and we found no membership attack designed for steering vectors in multimodal in-context learning. Algorithm 1 is therefore an attack we designed for this setting, not one taken off the shelf. It is also the stronger choice: our white-box attacker, with access to the activations, is strictly stronger than an output-only one, so the bound we show against it already covers the output-only threat model.

A few cells still show high accuracy at a small ε label because Figure 3's labels used the old, smaller sensitivity, so each column really sits at a larger budget than its label says; once relabeled, these cells fall at budgets where the attack is expected to still succeed. The remaining mid-range cells each come from only 20 targets, so they cannot be told apart from 50%. We will correct Section 3 and Algorithm 1 to describe the two vectors the attacker compares, and Figure 3 will use the corrected budgets with 100 targets per cell.

We hope these responses address the reviewer's concerns.

Reviewer_bNEEOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies privacy issues in multimodal ICL. The authors identify a critical privacy vulnerability in one such method, MTV, by proposing a simple yet effective membership inference attack and demonstrating its empirical success. They then privatize MTV drawing the classical DP techniques, addressing the issue. Despite the privacy guarantees, the utility remains quite good in the moderate privacy regime. The experiments are fairly comprehensive, and the results are mostly convincing.

Critical strengths

The paper tackles a timely topic. The motivation is clear, as demonstrated by the membership inference attack results. The proposed fix is fairly straightforward yet effective, as shown by its empirical success in the moderate-to-strong privacy regime.

Critical weaknesses

See below.

Questions for authors
  1. If i'm reading it correctly, the main experiment (Tabs 2 and 3) set S\mathcal{S} to be the set of all layers? Since the number of layers is usually not huge (for small models) and its contribution to the sensitivity is only via a square root factor, shall we always use S\mathcal{S} = all layers?

  2. It seems that the number of demonstrations does not affect the privacy at all (as long as we have the fixed number of chunks mm), which makes sense to me. And in fact, we can increase the number of demonstration and mm in the same time, which would make the sensitivity even smaller. So maybe your method is even better in this regime? Also, is there a reason not to set K=1K = 1 or at least as small as possible? cuz then for a given number of total examples, you have larger mm and thus smaller sensitivity. I thought that maybe KK cannot be too small to make vanilla MTV work but I had a hard time interpreting Tab 7. Could you please explain it a bit more?

  3. The explanation of "DP-MTV consistently exceeds non-private MTV" does not really convince me. To really explain what's going on, it would help to run a clip-only-no-noise and a no-clip-and-only-noise variants. I think this may even shed some light to the public MTV methods and would strengthen the paper.

  4. For the clipping parameters, I can't tell exactly what values of CC's were tried in Fig 5, but from the figure, it seems that you tried 1e-1, 1, 2, ...? In some DP-SGD papers, people suggest that it's best to clip 100% of the per-sample gradients, and I wonder if this is also true (for clipping the activations) in your case as well.

3 comments

We thank the reviewer for finding the motivation clear and the fix effective in the moderate-to-strong privacy regime. Point-by-point response below. This comment covers Q1 to Q3, and a second comment covers Q4 and the sensitivity constant.

Q1: Since |S| enters the sensitivity only as a square root, should |S| always be all layers?

Yes, and our main experiments already use all 32 layers. The noise grows only with the square root of the number of layers, so four times as many layers adds only twice the noise. Using every layer is therefore cheap, and we do not limit |S|. We will state this choice and its reason in Section 5.

Q2: The number of demonstrations does not affect privacy, so why not use many small chunks, and how should Table 7 be read?

The reviewer is right: the guarantee depends only on the number of chunks m, not on the number of demonstrations K per chunk. So more private data lets us use more chunks at no extra privacy cost, and small chunks are a good choice. Table 7 (VizWiz, ε=1) supports this in two ways. First, holding K=8, accuracy stays between 48.7 and 50.6 as m rises from 50 to 1000, even though the sensitivity drops 20-fold; what changes is the spread across seeds, which tightens from about ±9 at small m to ±5 at large m, so more chunks mainly add stability. Second, holding m=100, K=2 matches K=8 (51.1 vs 52.0), so shrinking the chunks costs little accuracy. For the camera-ready, Table 7 will add K=1, where a chunk holds a single demonstration, and its caption will say what each row holds fixed.

Q3: The regularization explanation of why DP-MTV beats non-private MTV is unconvincing without clip-only and noise-only runs.

The gain over non-private MTV is on the classification tasks, which is where this ablation applies. To separate the three factors, we keep the selected heads fixed and vary only what is placed at them (Qwen-VL, public head selection, accuracy in %). MTV and DP-MTV use the same selected heads, so head selection is not what sets them apart; the difference is the vector placed there.

What is placed at the headsFlowersCUBDTD
nothing (zero-shot)545655
raw MTV vector797980
MTV vector, clipped, no noise959393
DP-MTV (clipped, plus noise)969393
ε=2 noise alone (no vector)959492
the same, but at random heads495050

The raw vector reaches 79 on Flowers, matching non-private MTV in Table 2 (77 and 78 there, averaged over three seeds), and all three tasks behave the same. The table reads off each factor in turn. Clipping the vector with no noise reaches DP-MTV's accuracy (95 vs 96), so clipping, not noise, produces the gain; adding the noise on top changes nothing (96 vs 95). The selected heads are also required: the same noise placed at random heads instead stays at chance. Noise and head selection are shared by MTV and DP-MTV, so the one step that differs, and that moves the number, is clipping.

Varying the clip size C alone, with no noise, makes this direct. As C grows and less is clipped, the clip-only accuracy falls smoothly back to raw MTV, which it reaches once C is large enough to clip almost nothing.

clip-only accuracyC=1C=8C=20C=100raw MTV
Flowers9593897979
CUB9393857979
DTD9393868080

Clipping is therefore the step that turns MTV into the stronger result, and relaxing it recovers plain MTV.

This gain is specific to the 2-way classification tasks, where non-private MTV is weak (77 to 78). On the VQA tasks, where MTV gives a real gain over zero-shot, DP-MTV matches rather than exceeds MTV, and the added noise has a real cost. For the camera-ready, this ablation will replace our explanation in Section 5, and we will add the matching VQA controls.

As the reviewer expected, this also says something about MTV itself. The clipped row is non-private MTV with its vector clipped, and on these tasks it gains 14 to 16 points over the raw vector (79 to 95 on Flowers, 79 to 93 on CUB).

Continued from the previous comment. This comment covers Q4 and the sensitivity constant.

Q4: Which clipping values does Figure 5 use, and is it best to clip nearly all activations, as in DP-SGD?

Figure 5 uses C = 0.5, 1, 2, 3, 5 and 10 at ε=1, and accuracy is highest at C = 0.5 to 1. At these small values almost every layer's activation is larger than C, so nearly all of them get clipped. In our clip-only runs on Flowers, all 32 layers are clipped at C=0.1, and 25 of 32 are still clipped at C=4. Clip-only accuracy is steady across this small-C range (95% at C=0.1 and C=1, 93% at C=8), so clipping almost everything costs nothing, as the DP-SGD advice the reviewer mentions suggests. Two separate effects pull accuracy down as C grows, and Figure 5 combines them. Without noise, a larger C clips less and the result slides back toward raw MTV (the clip-only sweep in Q3). With noise, the noise also grows with C. Figure 5 adds both, so its drop at large C is steeper than the clip-only slide alone. The Figure 5 caption will list these values.

A correction to the sensitivity constant (raised in review).

This was raised by other reviewers, and they are right: because removing a record re-forms the rest of its chunk, the correct sensitivity is a factor of two larger, Δ = 2√|S|·C/m. The more important question is whether this changes our results, and it does not.

A larger sensitivity means more noise at every privacy budget. We ran the public-data experiments on Qwen-VL with the noise set for the new bound, at the same budgets and with the same accounting as the submitted runs, so only the sensitivity changed. The trend is the same as in the paper, shifted to a slightly larger budget: only the useful range moves up, while the ordering of the tasks and which ones DP-MTV matches or beats stay as they were. DP-MTV still matches or beats non-private MTV on 5 of the 8 tasks: the three classification tasks (above 90% against 76 to 78 for MTV), VizWiz (55.5 vs 55) and PathVQA (28.5 vs 28) at ε=2, and VQA-RAD (41.9 vs 40) at ε=5. It falls short only on OK-VQA and TextVQA, where MTV itself gives little or no gain over zero-shot, as already seen in Table 2. At ε=1 the larger noise leaves too little signal on the VQA tasks, so we present ε=2 to 5 as the operating range.

For the camera-ready, Theorem 4.1, Algorithm 2 and Appendix B will use the new bound and state the full privacy accounting, including the random sampling of the private records it relies on. Tables 2 and 3 will report all three models with the new noise, averaged over three seeds as in the paper.

We hope these responses address the reviewer's concerns.

Thank you for the detailed response and the ablation is really helpful. It would be great to include the discussion of mtv vs. dp-mtv, as well as the comparison of clipping and noising, in the paper. It addresses all my questions and I would like to maintain my score.

Reviewer_sZc4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DP-MTV, a framework for differentially private many-shot multimodal in-context learning. It aggregates clipped activations from disjoint chunks of private demonstrations, adds Gaussian noise to construct private task vectors, and supports both public-data and private attention-head selection. Experiments across eight benchmarks and three VLM architectures demonstrate promising utility and resistance to a proposed membership inference attack.

Critical strengths
  1. The paper addresses an important problem of protecting sensitive multimodal demonstrations while enabling efficient in-context learning. 2.The framework is practically appealing, as private task vectors can be reused for unlimited inference queries without additional privacy expenditure. 3.The experiments cover diverse benchmarks and model architectures, with promising results for the public-data variant.
Critical weaknesses
  1. Theorem 4.1 claims an (\ell_2)-sensitivity of (\sqrt{|S|}C/m) under record-level add/remove adjacency. However, changing one record can change the activation of an entire chunk. Per-layer clipping alone only guarantees that the difference between two clipped chunk activations is at most (2\sqrt{|S|}C), rather than (\sqrt{|S|}C).

Furthermore, it is unclear how the fixed-size chunk partition is maintained across neighboring datasets without affecting multiple chunks. The privacy proof requires additional justification or corrected sensitivity accounting.

A similar concern arises in the private head-selection mechanism (Algorithm 3), where a fixed-size validation subset is sampled from private data. The claimed utility sensitivity of (C_{\mathrm{sel}}) requires justification under add/remove adjacency.

  1. Unclear membership inference attack formulation and validity.

Algorithm 1 takes the target task vector as input but does not use it in the decision rule. Instead, it compares an activation against independently constructed reference vectors. It is unclear how this procedure infers membership in the target private dataset, or whether itself depends on the private dataset.

Moreover, Figure 3 reports high attack success rates even at very small privacy budgets. If evaluated under a standard balanced neighboring-dataset membership game, these results would be difficult to reconcile with the claimed DP guarantees. The authors should clarify the threat model, the construction of member and non-member instances, and the relationship between the empirical attack and the formal privacy definition.

Questions for authors
  1. The framework builds on standard DP aggregation techniques applied to activation steering. Some prior work already studies differentially private activation steering and membership inference. The authors should better distinguish the methodological contribution beyond the multimodal extension.

  2. DP-MTV substantially outperforms non-private MTV on several classification benchmarks. Ablations separating the effects of clipping, Gaussian noise, and head selection are necessary to understand these improvements and ensure fair comparisons.

3 comments

We thank the reviewer for finding the problem important and the reusable private task vector practically appealing. Point-by-point response below. This comment covers Concerns 1 and 2, and two further comments cover Concerns 3 to 5.

Concern 1: The sensitivity in Theorem 4.1 should be 2√|S|·C/m, and the fixed chunks need justifying when a record is added or removed (W1).

The reviewer is right. The correct sensitivity is a factor of two larger than the bound in Theorem 4.1. Appendix B's bound holds only when the removed record is alone in its chunk. When the chunk holds other records, removing one changes the whole chunk's activation, and the old and new activations can point in opposite directions, so the change can reach twice the clipping bound. The sensitivity, which is the most one record can change the released vector and sets how much noise we add, is therefore Δ = 2√|S|·C/m, where |S| is the number of layers, C is the clipping size per layer, and m is the number of chunks. Our DP-SGD comparison in Appendix B does not apply, because DP-SGD clips each record separately, while a chunk combines several records.

The chunks are disjoint and fixed before looking at the data: each record belongs to exactly one chunk, and the number and size of the chunks do not depend on which records are present. Adding or removing one record therefore changes only its own chunk and leaves every other chunk unchanged, which limits the sensitivity to a single chunk. That one chunk's change is two-sided, because removing the record still leaves the rest of its chunk, which re-forms into a different clipped activation; this is the factor of two above. We will make the chunk placement explicit in Theorem 4.1 and Section 4.

A larger sensitivity means more noise at every privacy budget. We ran the public-data experiments on Qwen-VL with the noise set for the new bound, at the same budgets and with the same accounting as the submitted runs, so only the sensitivity changed. The trend is the same as in the paper, shifted to a slightly larger budget: DP-MTV still matches or beats non-private MTV on the same 5 of the 8 tasks (table below), and the same ordering of tasks holds. It falls short only on OK-VQA and TextVQA, where MTV itself gives little or no gain over zero-shot, as already seen in Table 2. At ε=1 the larger noise leaves too little signal on the VQA tasks, so we present ε=2 to 5 as the operating range.

Qwen-VL, public datazero-shotnon-private MTVDP-MTV ε=2DP-MTV ε=5
VizWiz355555.560.3
PathVQA252828.529.1
VQA-RAD38400.341.9
OK-VQA586043.046.4
TextVQA686447.344.9
Flowers547798.093.8
CUB577891.693.8
DTD547692.793.1

The zero-shot and non-private MTV columns are the Qwen-VL rows of Table 2.

For the camera-ready, Theorem 4.1, Algorithm 2 and Appendix B will use the new bound and state the full privacy accounting, including the random sampling of the private records it relies on. Tables 2 and 3 and Figure 3 will report all three models with the new noise, averaged over three seeds as in the paper.

Concern 2: The fixed-size validation set in Algorithm 3 may break the C_sel bound (W1).

Algorithm 3 has two private steps, and each stays private when a record is added or removed.

The first step releases θ, the average of each chunk's p head probabilities over c chunks. Each probability is between 0 and 1, so the biggest change one record can make is to move all p of them from 0 to 1. Its sensitivity √p/c already uses this worst case, so it covers any change to a chunk and needs no factor of two.

The second step picks the final mask with the exponential mechanism, scored on validation records drawn at random. These records are drawn without looking at their content. One record then changes at most one validation loss, and each loss is capped at C_sel, so one record moves a candidate's score by at most C_sel. This gives (ε_mask, 0)-DP.

The C_sel bound holds for the results we report. The validation set is drawn without looking at the private records, so adding or removing one record changes at most one score, whatever the dataset size. The reviewer is right that a cleaner way to state the same guarantee is to include each record independently with a fixed probability; this is only a restatement of the construction, not a change to the mechanism or to the numbers above, and we will adopt it in the camera-ready to make the C_sel bound self-evident.

Continued from the previous comment. This comment covers Concerns 3 and 4.

Concern 3: Algorithm 1 does not use the released vector, and the attack succeeds at small ε (W2).

The released vector is in fact what the attack uses; Section 3 stated this unclearly. The attack is the standard membership test that differential privacy limits. The attacker is given two task vectors, both released by DP-MTV with independent noise: one built from demonstrations that include the target x* (τ_mem), one from the same demonstrations with x* replaced by another record (τ_non). It guesses that x* is in whichever vector is more similar to x*'s own activation a*. The decision is made on the released vectors themselves, so they are central to the attack, not unused. Section 3 described the attacker as building the vectors from background data and listed an input τ that the decision never uses, which made it look as if the released vector played no part; we will correct Section 3 and Algorithm 1 accordingly.

Exactly one vector contains x*, so a random guess is right half the time, and differential privacy drives the attacker's accuracy toward that 50% as ε shrinks, which is the trend in Figure 3. A few cells still show high accuracy at a small ε label because Figure 3's labels used the old, smaller sensitivity, so each column really sits at a larger budget than its label says; once relabeled, those cells fall at budgets where the attack is expected to still succeed. The remaining mid-range cells each come from only 20 targets, so they cannot be told apart from 50%. Figure 3 will use the corrected budgets with 100 targets per cell.

Our attacker is also stronger than one in deployment, not weaker: it sees the activations and uses the public model, so anyone holding the released vector and the public model can reproduce every output the steered model gives. The output-only attacker in Section 4 is therefore a weaker special case of ours, and defending against the stronger one implies defending against it. The paper's limitations already note this white-box assumption.

Concern 4: DP-MTV may only extend existing private steering and membership attacks to images (Q1).

The closest earlier work is PSA [1], which releases a private steering direction, built from pairs of contrasting text examples, to control alignment behaviors, and tests it with a membership attack. The other is DP-ICL [2], which makes each answer of few-shot prompting private. DP-MTV differs from both in four ways. First, it makes the many-shot in-context signal private, the average attention-head activations over hundreds of demonstrations, and it replaces activations at selected heads instead of adding a direction. Second, which heads to use is also learned from the private data, so we provide a private way to choose them. Third, unlike DP-ICL, which pays privacy budget for every query it answers, DP-MTV pays once for the released vector, which matters most at small budgets. Fourth, the membership attack is a further contribution, not a reused one: we found no membership attack for steering vectors in multimodal in-context learning, so Algorithm 1 is one we designed for this setting, rather than an existing attack extended to images.

We tested this third point on DP-ICL's own text benchmarks. Because DP-ICL's original code needs an OpenAI model that is no longer available, we reproduced its method, 100 prompts with 4 different examples each, combined by a noisy majority vote over 500 test questions. DP-MTV uses 100 chunks of 4 examples each. Both methods use Qwen2-7B-Instruct, δ = 10⁻⁵, and no subsampling; the camera-ready will add seeds.

Accuracy (%)ε=1ε=2ε=5
TREC, DP-ICL212538
TREC, DP-MTV807780
AG-News, DP-ICL517285
AG-News, DP-MTV788470

DP-ICL spreads its budget over all 500 questions, so on the six-class TREC task it stays far below the 87% of the same prompts without privacy at every budget, while DP-MTV stays within 10 points of that 87%. The gap grows with the number of questions, because DP-ICL pays for each one. On the easier AG-News, both methods do well and DP-ICL is close at ε=5, once its noise per question is small. For the camera-ready, we will add PSA to the related work, this comparison to the appendix, and seeds for both methods.

Refs.

[1] Goel, A., Hu, Y., Gurevych, I., & Sanyal, A. (2025, May). Differentially private steering for large language model alignment. In International Conference on Learning Representations (Vol. 2025, pp. 27723-27744).

[2] Wu, T., Panda, A., Wang, J. T., & Mittal, P. (2024, May). Privacy-preserving in-context learning for large language models. In International Conference on Learning Representations (Vol. 2024, pp. 20005-20040).

Continued from the previous comment. This comment covers Concern 5.

Concern 5: The gain over non-private MTV needs an ablation separating clipping, noise and head selection (Q2).

The gain over non-private MTV is on the classification tasks, which is where this ablation applies. To separate the three factors, we keep the selected heads fixed and vary only what is placed at them (Qwen-VL, public head selection, accuracy in %). MTV and DP-MTV use the same selected heads, so head selection is not what sets them apart; the difference is the vector placed there.

What is placed at the headsFlowersCUBDTD
nothing (zero-shot)545655
raw MTV vector797980
MTV vector, clipped, no noise959393
DP-MTV (clipped, plus noise)969393
ε=2 noise alone (no vector)959492
the same, but at random heads495050

The raw vector reaches 79 on Flowers, matching non-private MTV in Table 2 (77 and 78 there, averaged over three seeds), and all three tasks behave the same. The table reads off each factor in turn. Clipping the vector with no noise reaches DP-MTV's accuracy (95 vs 96), so clipping, not noise, produces the gain; adding the noise on top changes nothing (96 vs 95). The selected heads are also required: the same noise placed at random heads instead stays at chance. Noise and head selection are shared by MTV and DP-MTV, so the one step that differs, and that moves the number, is clipping.

Varying the clip size C alone, with no noise, makes this direct. As C grows and less is clipped, the clip-only accuracy falls smoothly back to raw MTV, which it reaches once C is large enough to clip almost nothing.

clip-only accuracyC=1C=8C=20C=100raw MTV
Flowers9593897979
CUB9393857979
DTD9393868080

Clipping is therefore the step that turns MTV into the stronger result, and relaxing it recovers plain MTV.

This gain is specific to the 2-way classification tasks, where non-private MTV is weak (77 to 78). On the VQA tasks, where MTV gives a real gain over zero-shot, DP-MTV matches rather than exceeds MTV, and the added noise has a real cost. For the camera-ready, this ablation will replace our explanation in Section 5, and we will add the matching VQA controls.

We hope these responses address the reviewer's concerns.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.