acceptodds
Under review as a conference paper at ICLR 2027

No TPU Left Behind: Retrofitting Side-Channel Protection into Edge TPUs

Abstract

Side-channel attacks can recover neural network parameters from physical signals, even on commercial edge accelerators. Existing defenses require changes to hardware, instruction set, or compiler, and cannot be deployed on fixed-function platforms such as TPUs. We present the first training-time defense that protects models on off-the-shelf TPUs without modifying the hardware or firmware. Our approach trains multiple functionally equivalent parameter versions per layer and randomly composes them at inference. This reduces the correlation that side-channel attacks rely on while preserving model accuracy. We enforce diversity between parameter versions by adding a regularization term in the loss function during training. We show that this diversity increases leakage variance while leaving the mean signal unchanged, which provably reduces the signal-to-noise ratio exploited by attackers. We derive theoretical bounds that relate leakage to the number of parameter versions and their pairwise distance, and provide a simple calibration method to predict leakage for new configurations before deployment or side-channel measurements. We implement our method on a Google Edge TPU and evaluate it on representative and real-world models. Our defense, in a high-diversity configuration, can hide leakage by reducing the Test Vector Leakage Assessment t-score below the standard leakage detection threshold of 4.5 with less than 1% accuracy change and moderate overhead. Our results thus show, for the first time, that training-time defenses can provide practical side-channel protection for widely deployed AI hardware.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Side-channel, model parameters, shuffling, defense, parameter stealing attack, Edge TPU

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_UuM9OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this work, a methodology for defending neural network weights against side-channel attacks is proposed. The methodology is intuitive and effective and trains alternative versions of layers for each part of the network to limit the leakage of individual paramaters/paths. Authors show this effectively mitigates first-order SCA leakage, and provide some initial results indicating

Critical strengths
  • Important problem. I think SCA defenses for ML-models are useful and more practical approaches that work on standard hardware are especially relevant.
  • Intuitive and simple approach that works well to mitigate direct leakage of weights with relatively minimal classification performance impact.
Critical weaknesses
  • The adaptive attack procedure does not seem strong enough. I would suggest considering a profiled setting where the attacker can train a small classifier model directly on traces to predict r1r_1. While this is less realistic as a threat model, it ensures we see what is possible against this defense. I would not really consider failing against this threat model to be a weakness, however, I am not convinced by the current unsupervised approach as it depends on getting each step right which is hard to do and potentially overly optimistic. The profiled setting would show whether an attack is feasible in worst-case settings, which also gives broader insights into what might be possible for a more determined unsupervised attacker.
  • The overhead is unclear to me (both training and inference). During training what is the memory overhead and extra time cost? Similarly `for each inference' might mean per batch of inputs or per single input. Additionally, if each batch uses a fixed path, could an adaptive attacker not exploit this? Conversely, if you have to swap path per input I imagine inference overhead is quite high?
  • Only CNNs/MLPs are tested. Maybe a small ResNet is useful for generality of the approach to more modern architectures.
Questions for authors
  • Can you run an adaptive attack where you try to predict what version of each/a layer is used using a supervised classifier (if traces are synchronized a small MLP should be reasonably quick to train)?
  • Can you provide overhead numbers for both training/inference n terms of memory footprint and time?

minor clarifications:

  • Is it really necessary to already have copies do this during training? Is it not feasible to take a trained, unprotected model and create alternative new versions of layers that preserve the input-output behavior per layer?
  • How does this combine with cryptanalytic extraction? It seems like it should make it harder implicitly? Can this defense be combined with [40]?
2 comments

We thank the reviewer for their valuable feedback and thoughtful questions. We appreciate the time and effort taken to review our submission. Below, we address weakness (W) and questions (Q) in detail.

W1/Q1. We implement a supervised deep-learning side-channel analysis that predicts the parameter-version selector r1 using 1D CNN and ResNet architectures trained on aligned EM traces of DENSE_1, with 2,000 fixed-input traces (for the lack of time to train on 20K raw measurements) per selector and a 60/20/20 split. For V=3, both achieve ~99% test accuracy against a 33.33% baseline; for V=8, ~99.5% against a 12.50% baseline. These results are obtained under three favorable conditions not available to our threat model attacker: (i) the execution region is pre-identified, (ii) traces are pre-aligned, and (iii) ground-truth selector labels are available for supervised training, obtained in our experiments from instrumented acquisition metadata. Under practical threat model, the attacker cannot observe or control selector labels, so obtaining equivalent labeled traces requires additional access beyond what is assumed. The unsupervised attack reflects what is feasible without labeled data, while the profiled attack shows the worst-case ceiling. Additionally, successful selector classification does not directly recover model parameters, though it may facilitate subsequent extraction.

W2 Training memory overhead: Parameter memory scales exactly V× since V independent weight matrices are stored per protected layer, reflected in the model size column of Table 1 (2.29–4.33× for MNIST MLP at V=2–8). Gradient memory is effectively 1× per layer regardless of V, since Algorithm 1 zeros out gradients for all versions not on the active path. Only one version accumulates gradients per mini-batch. Training time overhead: The increase (1.85×–30.10×) comes from two sources: the forward/backward pass through one randomly selected path per mini-batch, comparable to single-model training; and the diversity loss pairwise similarity computation across all V versions, which scales as O(V²) per layer and dominates at large V, explaining the superlinear growth in Table 1. Per-input vs per-batch path selection: Path selection happens per inference call, where each call processes a single input. The selector vector is sampled freshly per input not fixed across a batch. So each TFLite interpreter call receives its own independently sampled selector tensor. The inference time in Table 1 (averaged over 100 runs) reflects this per-input overhead.

W3 We evaluate EfficientNet real-world model attacked in prior work [32] and report security improvement in Table I.

Q2 Please refer to Table 1 where we report model size, training and inference time and security improvement for MLP, CNN and EfficientNet models.

Q3 While generating diverse versions post-hoc from a pre-trained model is appealing, it faces a fundamental challenge: composability. The defense requires that any combination of versions across layers produces accurate output, not merely that each version is accurate in isolation. Post-hoc perturbations to layer i change its output distribution in ways that layer i+1 trained on a fixed predecessor was never exposed to, breaking composability. Additionally, perturbations large enough to provide meaningful Euclidean distance for TVLA suppression are likely to degrade accuracy, while perturbations small enough to preserve accuracy may not provide sufficient diversity to suppress leakage below 4.5. The proposed training procedure solves both problems by construction: stochastic path selection guarantees composability across all V^L combinations, and the diversity regularizer λdiv provides principled control over the diversity-accuracy trade-off.

Q4. The proposed defense implicitly hardens against cryptanalytic extraction because these attacks rely on deterministic input-output behavior. The same query must produce the same response for the algebraic recovery to work. Since our defense randomly selects a different version combination per inference, the same input produces different outputs across calls, breaking this consistency requirement and increasing the query complexity of any cryptanalytic attack. Furthermore, the proposed defense and [40] are directly combinable, as both operate at training time through loss modifications targeting orthogonal properties. Cryptanalytic defense enforces query-response obfuscation against black-box API attacks while our diversity regularizer enforces parameter separation for side-channel resistance. The two terms can be added into a single unified training objective, providing simultaneous protection against both physical side-channel extraction on edge devices and cryptanalytic extraction in MLaaS settings.

Reviewer_UuM9OpenReview ↗

Response

I would like to thank the authors for their rebuttal. After also reading the other reviews and answers, as remarked by reviewer PKWq I think a core concern that complicates the evaluations is the lack of a concrete attack (which works on the unprotected model to retrieve weights) to measure defense performance on. Overall, I am therefore not convinced the parent in its current form clearly beats the bar for acceptance. I will reduce my score to 2. I think a lot of the issues that the paper is addressing are real and I like the simplicity of the approach. However, in addition to the remaining questions on my original review (see below), the current manuscript does not clearly demonstrate that the defense actually helps against parameter retrieval attacks that would otherwise succeed.

W1/Q1: Besides this, the adaptive results with supervised models recovering it are not helpful to the story in my view. While indeed the assumptions are quite strong, (i) and (ii) should be feasible to waive for attackers that are optimizing DL models (for the attack, very long misaligned raw traces are feasible to use for attack even against masked cryptographic traces, see e.g. [R1,R2]). If accuracy is 99% with these assumptions indicating that the value is index is leaking quite strongly, I would expect that a determined practical attacker could feasibly overcome these limitations.

W2/3: Thanks for clarifying, sorry I missed these before.

Q2: On the inference overhead in Table 1: I saw this, my question was more whether this is per image overhead, or overhead for processing 100 images sequentially (incl. per image weight swaps for the protected case)? I think both should be included.

Q3: While I appreciate these issues could/would be real in theory, authors have not demonstrated them empirically. Why would a clever post-hoc procedure not be able to overcome these issues? Maybe through using your objectives only for post-hoc construction of alternative layers? I would recommend including attempting some naive post-hoc baselines that demonstrate these issues.

Q4: I would again recommend demonstrating this experimentally. In my view, in a threat model where we care about side-channel extraction we should likely care at least as much about cryptanalytic extraction. Therefore, some experiments here testing how this defense affects the resistance against these would be useful. Additionally, in case it does not resolve the attack, the combination with [40] should also be relatively straightforward to test.

[R1]: Lu, X., Zhang, C., Cao, P., Gu, D., & Lu, H. (2021). Pay attention to raw traces: A deep learning architecture for end-to-end profiling attacks. IACR transactions on cryptographic hardware and embedded systems, 235-274.

[R2]: Perin, G., Wu, L., & Picek, S. (2022). Exploring feature selection scenarios for deep learning-based side-channel analysis. IACR Transactions on Cryptographic Hardware and Embedded Systems, 828-861

Reviewer_3BrCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper trains each protected layer with parameter versions and randomly selects one per inference on a Coral Edge TPU, requiring no hardware changes. Dense layers use weight stacking with a one-hot selector; convolution layers compute all convolutions and select one output. A TVLA-based analysis relates leakage to version count and inter-version distance. The evaluation reports TVLA reductions and accuracy/cost overheads on MNIST, CIFAR-10, and two EfficientNet layers.

Critical strengths
  1. The problem is practically relevant. Developing defenses for deployed accelerators whose hardware and firmware cannot be modified addresses an important deployment limitation of existing protection mechanisms.
  2. The work includes implementation and measurements on a real platform. EM experiments on Google Coral, together with accuracy, latency, and model-size measurements, provide useful evidence for assessing the engineering feasibility of the approach.
Critical weaknesses
  1. The defense design and its deployment cost are insufficiently justified. Composability also reduces the attacker's recovery requirements: under the claimed composability property, recovering any one complete version of each protected layer suffices to assemble a usable model. The attacker need not recover all versions or network configurations. Consequently, the existence of V^L combinations does not itself establish a corresponding increase in attack difficulty. Meanwhile, the MNIST models require 2.29–4.33× the baseline model size and 3.43–5.57× the inference latency; the smaller size increase for EfficientNet covers only its first two layers. For memory- and latency-constrained edge inference, I do not consider these substantial costs acceptable on the basis of the demonstrated security benefit. The non-monotonic benefits in Table 1, such as the lower maximum improvement at V=7 than at V=6, further weaken the justification of the cost–benefit trade-off.
  2. The theory and evaluation do not substantiate the claimed security gains. The appendix treats a variance upper bound obtained through Cauchy–Schwarz as a guarantee of variance growth with parameter distance, then derives a t-statistic upper bound. This inference is not valid: greater parameter distance need not yield greater leakage differences along actual input directions. The evaluation lacks ablations that isolate stochastic path selection, diversity regularization, and additional computation, leaving the source of TVLA reduction unclear. It also reports neither actual parameter recovery nor the performance of an extracted model. Finally, subtracting cosine similarity in Algorithm 1 is inconsistent with the stated goal of discouraging similarity and requires reconciliation with the implementation.
  3. Several conclusions require configuration-specific evidence or narrower wording. Figure 2 reports below-threshold leakage for CIFAR-10's DENSE_2 region at 20K total traces using V_conv=2, V_dense=8, and lambda_div=100. Table 1 instead reports CIFAR-10 with V=2 throughout and lambda_div=60. Accuracy, training time, inference latency, and model size are not reported for the exact Figure 2 configuration, leaving the abstract's joint claim of below-threshold leakage, less than 1% accuracy change, and moderate overhead without matching measurements. This is a missing-evidence issue, not a direct contradiction between the abstract and the 1.3-percentage-point loss of a different configuration. The claim in Appendix A.5.2 of stronger suppression in DENSE_2 also needs qualification: at the rightmost plotted point in Figure 8, V=2 reduces DENSE_1 from 279.3 to 136.8 (about 51%), but DENSE_2 from 138.88 to 91.23 (about 34%). DENSE_2 retains a lower absolute t-statistic, but its reduction relative to its own baseline is smaller in this setting. Finally, Section 6 mentions a reduction from a baseline above 500 without identifying a corresponding figure, table, or acquisition configuration.
Questions for authors

The first two would most affect my rating:

  1. Can you run a defense-aware, layer-wise extraction attack (e.g. CPA or template) that recovers one version per layer and assembles them into a single network, and report recovered-weight error and the assembled model's accuracy versus trace count? Given composability, why would this model not reach near-baseline performance?
  2. For convolution layers, does the order of the V executions actually change across inferences in the compiled graph? Please evaluate segmentation/re-alignment attacks and compare against a latency-matched control using identical versions or dummy computations to separate parameter diversity from desynchronization.

Some others:

  1. Section 3.2 says the hardware generates the one-hot selector, whereas Appendix A.4.2 says the compiled model takes one selector tensor per protected layer. Which component samples and supplies these tensors at inference, and can an attacker observe or control them?
  2. Eq. (37) only upper-bounds the version-induced variance and therefore does not provide the lower bound needed for Theorem 7's t-score upper bound. Moreover, Figure 6 appears inconsistent with Eq. (4): the plotted dmin⁡d_{\min} values give nearly equal V−1Vdmin⁡2\frac{V-1}{V}d_{\min}^2 at V=2V=2 and V=8V=8, yet the predicted t-score drops from approximately 46 to 16.5. Please clarify this calculation and why KK can be shared across separately trained models.
  3. Please reconcile the reporting gaps: 3.81–48.04× (intro) vs 1.48–48.04× (Table 1); the "stronger DENSE_2 suppression" claim (holds at V=8V=8 but not V=2V=2); and the Figure 2 below-threshold configuration's missing accuracy/cost numbers.
1 comment

We thank the reviewer for their valuable feedback and thoughtful questions. We appreciate the time and effort taken to review our submission. Below, we address weakness (W) and questions (Q) in detail.

W1. Yes, recovering one complete version per layer suffices. However, since the active version per layer is selected at random, the probability that a trace uses a specific version across all L layers is (1/V)^L, requiring V^L times more measurements. Section 5.3 demonstrates that the attacker cannot determine the active version on a trace, making this collection cost unavoidable in practice. On deployment overhead: The overhead figures correspond to highest-V all-layer configurations. The defense is designed for selective application to the most vulnerable layers. All alternative defenses report 3–50× resource overheads and cannot be deployed on Edge TPU at all. On non-monotonic security improvement: As V increases, dmin decreases because additional versions have less freedom to remain separated while preserving accuracy. Since leakage reduction depends jointly on V and dmin, a decrease in dmin can offset the variance gain from additional versions, producing non-monotonic behavior. The calibrated prediction model in Section 4.3 accounts for this, enabling designers to identify configurations where increasing V improves security before deployment.

W2/Q4. In order to upper bound the t-score, one must lower bound the denominator. We acknowledge that Eq. (24) applies in the wrong direction, and clarify the fix here. The inner product satisfies ⟨ΔW, xf⟩² ≥ c·‖ΔW‖² for some c > 0, given that xf is not orthogonal to the ΔW. The ⟨ΔW, xf⟩ is nonzero, since if ⟨ΔW, xf⟩ = 0, then Wa and Wb would produce identical outputs on xf. The regularization term ensures the weight versions produce different outputs on real inputs. This gives the lower bound, and the chain of reasoning follows: (La − Lb)² ≈ α²⟨ΔW, xf⟩² ≥ α²·c·‖ΔW‖² → σ²_f = (La − Lb)²/4 + σ²_ε ≥ α²c‖ΔW‖²/4 + σ²_ε → Lower bound on denominator: √(σ²_f/Nf + σ²_r/Nr) ≥ √(C·‖ΔW‖² + C₀) → t ≤ K / √(C·‖ΔW‖² + C₀) This establishes Theorem 7 correctly: as ‖ΔW‖ increases, the upper bound on t decreases, meaning increasing parameter diversity reduces the ceiling on leakage. We will correct this.

W3/Q5. The strongest leakage suppression is demonstrated on the CIFAR-10 CNN with λdiv=100. The overhead and security for this configuration are: Accuracy: 93.2% (−0.72pp), Training time: 18.6s (7.7×), Inference time: 22.001ms (9.91×), Model size: 8.4MB (7.6×), Security improvement for DENSE_2: 191.18×. We will add this analysis. Reduction from 500: This is observed on CNN model, where the 2nd convolutional layer of the unprotected model yields a t-score of 546, much higher than the baselines in prior works.

Q1. A layer-wise CPA or template attack requires traces where a single known parameter version is active but, the attacker cannot determine which version was active during any given trace as clustering achieves near-chance accuracy. Without this version-to-trace assignment, CPA must operate on a mixed trace set containing leakage from all V versions. Moreover, CPA is layer-wise, meaning errors propagate and compound across layers. This means full model assembly is not simply V independent recovery problems but a sequentially dependent chain where a failure at any layer corrupts all downstream recoveries. While composability means any recovered version combination produces a functional model, it does not reduce the difficulty of recovering any version accurately.

Q2. Yes, the execution order of the V convolutions change across inferences, confirmed through runtime prints that log the randomly sampled one-hot selector vector at each inference. The selector is generated at runtime on the host CPU and passed as an input tensor, so the compiled graph receives a different ordering signal at every inference while its structure remains static consistent with the Edge TPU’s constant-sized tensor requirement. On segmentation and re-alignment attacks: Without access to the selector vector, attacker cannot determine the correct segmentation without solving the clustering problem shown to be practically infeasible in Section 5.3. Guessing the ordering requires exhaustive search over V possible orderings per trace, exponentially increasing measurement cost.

Q3. The attacker has physical access to the device during inference but operates in a black-box setting without access to the host CPU’s state or the interface through which the selector is transmitted. The one-hot selector tensors are sampled by the host CPU at runtime by generating a random index per protected layer and passing it as an input tensor to the compiled TFLite model alongside the input. An attacker cannot recover the selector values (Section 5.3). Clustering attacks achieve only near-chance accuracy. The attacker therefore can neither observe nor control the selector tensors.

Reviewer_PKWqOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The author proposes a training-based defense against physical side-channel extraction of DNN parameters on commodity Edge TPUs. It trains multiple parameter versions per layer using stochastic path selection and diversity regularization, then randomizes version selection during inference. The authors derive a distance-based TVLA prediction model and evaluate EM leakage, accuracy, and overhead on MNIST, CIFAR10, and selected layers of EfficientNet-EdgeTPU-S. Security evaluation relies on first-order fixed vs. random TVLA and parameter version clustering experiments.

Critical strengths
  1. The author try to addresse an important practical problem: retrofitting model-confidentiality defenses onto commodity accelerators without hardware or firmware changes.
  2. The stochastic training strategy enables layer wise parameter variation while preserving near baseline accuracy in the reported configurations.
  3. Physical EM measurements on a commercial Edge TPU demonstrate substantial reductions in first order leakage detectability for some evaluated layers.
Critical weaknesses
  1. The theoretical upper bound is unsupported because the derivation upper-bounds leakage variance, whereas upper-bounding requires a variance lower bound.
  2. The security evaluation relies on first-order TVLA and narrowly restricted clustering without demonstrating increased resistance to actual parameter extraction attacks.
  3. Figure 6’s predicted t-statistics appear numerically inconsistent with Eq. (4) and the parameter distances displayed alongside them.
  4. Accuracy and overhead are not reported for the exact configurations demonstrating the strongest leakage suppression, leaving the claimed practical trade-off unsubstantiated.
Questions for authors
  1. What variance lower bound establishes Theorem 7, given that Eqs. (24) and (37) provide upper bounds?
  2. What are the parameter recovery success rates and measurement costs for protected and unprotected models under the same adaptive extraction attack?
  3. What prevents extraction from individual convolutional branches when every parameter version executes during each inference?
  4. What calibration constants and distance values reproduce Figure 6s predictions from Eq. (4)?
  5. What are the deployed accuracy, latency, model size, and all-layer leakage results for the exact strongest-suppression configurations?
2 comments

We thank the reviewer for their valuable feedback and thoughtful questions. We appreciate the time and effort taken to review our submission. Below, we address weakness (W) and questions (Q) in detail.

W1/Q1. We thank the reviewer for this careful observation. Yes, in order to upper bound the t-score, one must lower bound the denominator. We acknowledge that Eq. (24) applies in the wrong direction, and clarify the fix here. The inner product satisfies ⟨ΔW, xf⟩² ≥ c·‖ΔW‖² for some c > 0, given that xf is not orthogonal to the ΔW. The inner product ⟨ΔW, xf⟩ is nonzero for the fixed TVLA input xf, since if ⟨ΔW, xf⟩ = 0, then Wa and Wb would produce identical outputs on xf. The regularization term ensures the weight versions produce different outputs on real inputs. This gives the lower bound, and the chain of reasoning follows: (La − Lb)² ≈ α²⟨ΔW, xf⟩² ≥ α²·c·‖ΔW‖² → σ²_f = (La − Lb)²/4 + σ²_ε ≥ α²c‖ΔW‖²/4 + σ²_ε → Lower bound on the t-statistic denominator: √(σ²_f/Nf + σ²_r/Nr) ≥ √(C·‖ΔW‖² + C₀) → Upper bound on t: t ≤ K / √(C·‖ΔW‖² + C₀) This establishes Theorem 7 correctly: as ‖ΔW‖ increases, the upper bound on t decreases, meaning increasing parameter diversity reduces the ceiling on side-channel leakage. We will correct this.

W2. The choice of first-order TVLA as the evaluation metric is justified in Section 6 of the paper; we summarize the reasoning here for clarity. TVLA stands as the primary evaluation for end-to-end leakage analysis, as in certification frameworks like FIPS, ISO/IEC, and NIST. We choose TVLA over CPA/DPA because CPA/DPA is narrow and tests only specific parameter extractions (single weight recovery), TVLA is used to expose comprehensive leakage. Following standard practice across AI/ML attacks, we rely on TVLA as the sole evaluation metric (29, 2, 20, 6, 45). Higher-order TVLA is designed for masking-based countermeasures that split secrets into shares; our defense instead stochastically selects complete, functionally equivalent parameter values with no secret sharing, making higher-order TVLA inapplicable. The clustering analysis evaluates the adaptive attack to undo randomization before extraction. Section 5.3 demonstrates that version-specific trace separation is practically infeasible. Without this separation, any parameter extraction operates on a mixed signal from all V versions, making extraction difficult.

W3/Q4. We identified the source of the inconsistency. Figure 6(b) displays dmin values before quantization. The model is calibrated using the V=2 configuration of DENSE_2 with λdiv = 0.3, giving t1 = 138.88 (baseline), t2 = 46.0 (V=2 protected), and d = 483 (dmin at V=2). This yields: A/B = (2/483²) · [(138.88/46)² − 1] = 0.00006959. The dmin values used in the Eq. (4) predictions across V=2 to V=8 are [483, 487, 518, 564, 589, 619, 659] respectively. The dmin values increase with V, consistent with the larger parameter space of DENSE_2 providing greater freedom for versions to remain separated. We will correct this.

W4/Q5. The strongest leakage suppression is demonstrated on the CIFAR-10 CNN with λdiv=100. The overhead and security improvements for this configuration are: Accuracy: 93.2% (−0.72pp), Training time: 18.6s (7.7×), Inference time: 22.001ms (9.91×), Model size: 8.4MB (7.6×), Security improvement for DENSE_2: 191.18×. We will add this analysis to the revised manuscript.

Q2. Since each inference randomly selects one of V parameter versions per layer, an attacker collecting traces across inferences receives a mixture of signals from all V versions. As shown in Section 5.3, the adaptive clustering fails to separate traces by version. Because the attacker cannot identify which version was used during each inference, they cannot group traces version before attempting extraction. Even in the case where an attacker somehow identifies the active version, they would need to collect V times the number of measurements per layer, since only 1/V of all traces correspond to any given version. The paper reports security improvement as the factor increase in traces required for the protected model to reach the same t-score as the unprotected model (Table 1), which quantifies the increase in measurement cost. This ranges from 3.81× to 48.04× across evaluated configurations, meaning the attacker requires up to 48× more measurements to achieve the same leakage detectability as against the unprotected model.

Q3. We clarify that while all convolutions execute during inference, this does not enable individual branch extraction. As stated in Section 3.2, the V convolutions are executed in random order across inferences so that the attacker cannot assign a time window to a specific version across traces. Each trace has the V branches in a different order. This produces the same mixing effect shown for the dense layers, where clustering accuracies confirm that separating version leakage from the mixed signal fails.

Thank you for the detailed response. I went back through the manuscript with your answers in mind. I think protecting fixed-function accelerators without changing the hardware is worth pursuing. The question I keep coming back to is what the defense protects against on the Edge TPU.

  1. The defense is motivated by weight extraction on the Edge TPU, but as far as I can tell, no such attack is shown, either in the paper or in the cited work (W2/Q2). [41] recovers hyperparameters on the Edge TPU, [32] targets edge GPUs, and the systolic-array attacks in [66] and [43] target FPGA implementations. Recovering weights requires knowing where and when each weight is processed. As the paper notes, the Edge TPU floorplan is not public, and the paper does not characterize how weights are mapped onto the compute array or scheduled in time: the probe is fixed at a single package-level hotspot, and the traces are aligned to layer-level windows (Appendix A.4.1). TVLA on these windows detects data-dependent leakage, but it does not show that this leakage is sufficient to recover weights. Without a successful attack on the unprotected model, it is difficult to judge what the defense protects against. I would first demonstrate a weight-recovery attack on the unprotected model, then repeat the same attack with protection enabled. This would also answer Q2 directly.

  2. The configuration behind the sub-4.5 claim in Figure 2 is not in Table 1 (W4/Q5). Its inference and model-size overhead factors in W4 exceed those of every configuration in Table 1, and its accuracy (93.2%, −0.72 pp) is inconsistent with the 72.4% baseline there. Adding it to Table 1 with its baseline would show the actual trade-off.

  3. I could not reproduce Figure 6(a) from the constants in the response (W3/Q4). In addition, t1t_1 and t2t_2 appear to come from different trace budgets (Figure 8(b) vs. Figure 6), while Eq. (4) assumes a common budget, since AA and BB scale with 1/N1/N (Eqs. 42–43). The dmin⁡d_{\min} values used differ from Figure 6(b) as well. The predictions should be recomputed.

  4. Under the proposed fix to Theorem 7 (W1/Q1), cc depends on the direction of ΔW\Delta W relative to xfx_f, which dmin⁡d_{\min} does not capture. How then does A/BA/B calibrated at V=2V=2 carry over to other configurations?

  5. Since KK does not depend on VV or the version distances (Theorem 4) and tt grows with traces (Figure 8), the sub-4.5 result depends on the trace budget, as Section 5.2 notes; the finite 191.18×191.18\times improvement in W4 points the same way. The abstract should state this budget. Because the selector is an input tensor (Appendix A.4.2), a profiled attack on it should also be tested.

At present, the security claim rests on leakage detection rather than on resistance to a demonstrated attack. The new experiments and the revised derivation bear directly on the main claims, and that is more work than I would expect the authors to complete during the discussion period. I am keeping my score for this round, and I would be interested to see the work developed into a full revision for a fresh round of review.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.