acceptodds
Under review as a conference paper at ICLR 2027

TIDE: Defending Against Backdoor Data Poisoning via Dense Influence Subgraph Detection

Abstract

Targeted poisoning attacks perturb the training data of machine learning models to increase the probability that the trained models make certain errors. Stealthy backdoor attacks, which perturb data by injecting inconspicuous triggers into the features, pose a particularly serious threat to the reliability and security of machine learning. To detect and filter backdoor-poisoned data, we shift the focus from identifying individually anomalous data points to identifying groups of training data that have a disproportionately high impact on each other during training. Specifically, we first introduce an efficient approach for estimating how the inclusion of a data point in model training influences the output of the trained model for another data point. Leveraging the observation that this influence is disproportionally high between pairs of poisoned data, we introduce an efficient algorithm for detecting poisoned data as a dense subgraph in the graph formed by data points and the influence values between them. We evaluate our proposed approach against state-of-the-art backdoor poisoning attacks on benchmark datasets, demonstrating that it outperforms existing approaches for data cleaning in terms of reducing attack success and increasing model accuracy. We also demonstrate robustness against an adaptive attack that attempts to circumvent our defense by reducing the cross-influence of poisoned data.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Adversarial machine learning, Data poisoning, Backdoor attack, Poisoned data detection, Influence score

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

TIDE detects clean-label backdoor poisoning through collective training influence, without a known-clean reference set or known poisoning rate. It builds a source–probe influence graph from temporary training and uses greedy peeling to select suspicious samples before retraining. Across CIFAR-10, GTSRB, and additional Tiny ImageNet and VGG experiments, the authors report reduced attack success with generally preserved clean accuracy. For example, CIFAR-10 Narcissus ASR falls from 91.20% to 7.92%, although substantial residual ASR remains for the Wicked variants. The results support a useful cleaning trade-off; the paper interprets shared influence as the detection signal.

Critical strengths
  1. Cleaning clean-label poisons without trusted reference data is a meaningful problem: correct labels make suspicious samples harder to identify individually. Section 4's collective source–probe formulation is an interesting extension of existing influence analysis. Its novelty lies in the bipartite detection procedure, rather than introducing influence-based defense itself.

  2. Tables 1–2 and Appendix G provide reasonably broad evidence for mitigation across three datasets, two architectures, and five attack configurations. Comparisons include FLARE (2025), with PureGen (2024) in Appendix E. The results support a useful mitigation–accuracy trade-off, although the pairwise attribution and poisoning-specific interpretation require the separate checks discussed below.

Critical weaknesses
  1. Equations (4)–(5) distribute the batch's net logit-change magnitude across samples. Opposing sample effects can therefore cancel before attribution and receive zero scores. Cosine weighting also discards feature norms, so relative contributions can be distorted even with a frozen linear head. This limits the claimed pairwise attribution beyond the acknowledged difference from leave-one-out influence.

  2. Equation (7)'s poisoning specificity is unresolved. Higher poison-to-poison influence than the other pooled pair means does not ensure that the poisoned block outscores an equally sized benign cohesive block. Appendix A does not separate class or degree effects, and Appendix F's optimizer comparisons cannot validate that identification step.

Questions for authors
  1. Under what assumptions do Equations (3)–(5) approximate individual step effects? How are within-batch cancellation, feature norms, and momentum or backbone updates handled?

  2. What distinguishes the selected block from a clean, cohesive class or a high-degree source group? Can class-matched clean blocks and marginal rankings be compared using the same influence matrix?

  3. At low poisoning rates, how does proportional peeling behave when the probe contains no poison? Please clarify the final partition-aggregation rule and Table 10's stated 0.5% versus implied 5% removal budget.

4 comments

We thank you for your thoughtful review.

Assumptions, momentum, and backbone updates. At each step, we measure the actual change of the probe logits on the complete model (Eq. (2)), so backbone updates are included in the quantity being attributed. Eqs. (3)–(5) then distribute this change among the batch samples using a last-layer, first-order approximation. Each sample’s share is proportional to the similarity of its last-layer feature to the probe’s feature, weighted by its prediction error. We therefore present it as an approximation, not as leave-one-out influence. Since the tracking run uses SGD with momentum (Section 5.1), the measured change also contains a carry-over from earlier batches, which our attribution assigns to the current batch. We regard this as part of the approximation.

Within-batch cancellation. The reviewer is right that the effects of samples in the same batch can cancel before attribution, so those samples receive zero (or very small) scores, since we distribute the batch’s net logit change. This is a limitation of a single step. However, batch composition is randomized across epochs and repetitions, so each sample’s accumulated score reflects its contributions across multiple batch contexts rather than a single batch. Samples that consistently influence the probe in many different batches tend to obtain a larger accumulated score (Eq. (6)), which is what the detection uses. Consistent with this, poison-to-poison pairs have the highest mean accumulated influence in our measurements (Section 5.5, Table 3).

Feature norms. We agree that cosine weighting discards feature norms, so relative contributions can be distorted. We use cosine similarity as a simple, scale-free weighting function. Prediction error contributes to each sample’s attribution weight, while the measured batch-level change determines the total magnitude to be allocated. Other similarity measures (e.g., unnormalized inner products) are possible, and we have added this to the limitations.

To examine the class-effect concern, we extend the analysis of Appendix A (now summarized in the new Section 5.5) by comparing poison-to-poison influence with clean-to-clean influence within the target class. Across all seven evaluated settings (Section 5.5, Table 3; full statistics in Appendix A, Table 4), poison-to-poison pairs have higher mean influence than clean-to-clean pairs both overall and within the target class. Target-class clean pairs also show higher mean influence than clean pairs overall, confirming that a class effect exists, which is why the within-class comparison matters. Nevertheless, the higher mean influence of poison-to-poison pairs persists when the clean comparison is restricted to the target class.

Table R3. Mean (median) pairwise influence; Clean denotes all clean samples, and Clean_t denotes clean samples from the target class.

DataAttackPoison →\to PoisonClean →\to CleanClean_t →\to Clean_t
CIFAR-10Narcissus0.1193 (0.0109)0.0053 (0.0026)0.0476 (0.0231)
CIFAR-10COMBAT0.6100 (0.5746)0.3236 (0.1895)0.4380 (0.1239)
CIFAR-10Wicked-BadNet0.7136 (0.1299)0.2553 (0.0636)0.3655 (0.0893)
CIFAR-10Wicked-Blended0.6881 (0.1305)0.3686 (0.0897)0.4717 (0.2043)
CIFAR-10Wicked-SIG0.6767 (0.0976)0.2324 (0.1047)0.2444 (0.0568)
Tiny ImageNetNarcissus9.3556 (7.9551)7.3762 (5.4336)7.5548 (6.1562)
Tiny ImageNetWicked-BadNet8.5550 (5.8516)7.5446 (5.4785)7.5703 (6.2266)

Proportional peeling when the probe contains no poison. Proportional peeling removes samples from the side with the smaller removed fraction, which assumes that poisons are distributed over the source and probe sets, as expected from a random partition. If the probe contains no poison, this assumption does not hold and peeling on the probe side can only remove clean samples. To quantify this effect, we compared our setting with partitions in which all poisons are forced into the source set (so the probe contains no poison), using the same setting as our main experiments, averaged over three runs. With a random probe, TIDE achieves TPR 0.7260 and FPR 0.2597, and with a probe containing no poison, TPR 0.4045 and FPR 0.1744. Detection is therefore weaker when the probe contains no poison, but it does not collapse. We report this analysis in Appendix E.

Aggregation rule. In each repetition, the dataset is randomly partitioned into a source set and a probe set (the probe sets are disjoint across repetitions), and all samples are nodes of the bipartite graph. Greedy peeling yields a suspiciousness ranking over all samples, and the position krk_r in this ranking that maximizes the objective JJ defines the selected subset of repetition rr. We average the rank of each sample over the RR repetitions to obtain an aggregated ranking, and remove the top-ranked samples up to the average kˉ=1R∑rkr\bar k = \frac{1}{R}\sum_r k_r of the per-repetition cutoffs. We have added this rule to Section 4.2 (Multiple Random Partitions).

Table 10 (Table 13 in the revised paper). TIDE does not use a fixed budget. Greedy peeling returns the subsets that maximize the objective JJ. However, integer programming and simulated annealing solve a problem of a given size, so for a fair comparison between these approaches, we fixed the selection size to 5% of the training set for all of them when the poisoning rate is 0.5%. This fixed budget is used only in this comparison. We have corrected the stated selection size to 5% in Appendix G.

Thank you for the clarifications and additional analyses. The within-class comparisons and aggregation details are helpful. However, higher average poison-to-poison influence does not establish that the selection objective distinguishes poisoned blocks from equally sized cohesive clean blocks or high-degree groups. The acknowledged cancellation, feature-norm, and momentum effects also leave the reliability of pairwise attribution insufficiently established. I therefore maintain my original rating.

Reviewer_3qyDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors introduce TIDE - tracking influence for dense subgraph estimation to detect clean-label backdoor samples from a training set. The algorithm starts with randomly sampling the set into subgroups - source set and probe set. The source set’s influence over the probe set is measured as defined in Section 4. Backdoor sample detection happens over two stages. In stage 1, for every minibatch, the influence of source samples over probe samples is computed. Once this is done over several epochs, the directional graph is interpreted - with poisoned samples expected to encode a shared trigger-target association. The authors contend that poisoned source samples will exert high influence over poisoned probe samples, and create a region of dense high-influence across the two partitions. The authors use a greedy peeling heuristic to select poisoned samples which is their bipartite subgraph search algorithm. The authors lay out a threat model to empirically evaluate the efficacy of their data-cleaning algorithm in reducing attack success rates while sustaining high training accuracy across 5 attack types. They further compare their results with 5 defense baselines to benchmark against.

Critical strengths
  • The algorithm is simple and fast. It does not require clean validation data.
  • The bipartite approximation and random splitting into source and probe sub-groups makes the algorithm computationally trivial to implement.
  • The authors leverage the gradient alignment between source and probe samples that naturally follow clean-label attacks.
  • The authors consider a variety of clean label attacks and defenses in their experiment setup. TIDE seems to do a great job at preserving accuracy and reducing attack efficacy in these settings.
Critical weaknesses
  • In section 3, the authors propose a threat model where the defender does not have access to the poison ratio or an upper bound. The experiments across the paper use a ratio of 0.5% for CIFAR-10 and 0.03% to 0.3% for GTSRB, and yet the peeling budget seems fixed to 0.5%. It is unclear if a fixed peeling budget was chosen or not.

  • Baselines like Narcissus clean-label attack achieve high ASR with a very low attack budget. For instance their paper reports an 85% ASR with only 0.05% of samples poisoned in the tiny-imagenet dataset. With such low budgets, TIDE could possibly partition data into probe sets that may contain no poisoned samples. This paper misses a comment on the same.

  • Appendix A includes an important result. The main paper will benefit from talking more about the motivations and empirical evidence behind this paper.

  • Can the authors provide an ablation experiment to estimate the right number of epochs per dataset size - or signal their recommendation for others to follow ? It is unclear why 10 epochs were chosen, and if the signal needed for identifying poisoned samples can be captured before 10 epochs, making the runtimes possibly faster.

Questions for authors
  • This paper would greatly benefit from studying how (a) a varying poison rate affects this defense but more importantly (b) how the authors plan to tackle removal in those circumstances.

  • Can the authors address attacks under low poison ratios and explore how effective TIDE is to filter out poisoned samples under that setting ?

5 comments

Thank you for raising this point. TIDE does not use a fixed peeling budget. Its peeling and cutoff-selection procedures do not require the poisoning rate or an upper bound on it. As described in Section 4.2, greedy peeling tracks the objective value after each removal and returns the subsets that maximize it, so the size of the detected set is determined by the data. Across repetitions, we aggregate the resulting suspiciousness rankings.

The fixed selection size appears only in the comparison of subgraph search algorithms in Table 13 in Appendix G (Table 10 in Appendix F in the original submission). Our integer-programming and simulated-annealing formulations use a fixed selection size, so we ran all three algorithms with a 5% selection size for a fair comparison (note that the 0.5% selection size was a typo in our original submission). Thus, the selection size was never set to the true poisoning rate. This fixed size is not part of TIDE’s detection procedure. We have clarified this in Section 4.2 and Appendix G of the revised paper.

The number of poisons in a random probe depends on the poisoning rate ρ\rho and the probe size PP, with expected count ρP\rho P. Under a binomial approximation, the probability that the probe contains no poison is (1−ρ)P≈e−ρP(1-\rho)^P \approx e^{-\rho P}, which is about 0.7% at ρ=0.5\rho=0.5% and P=1,000P=1{,}000.

Our evaluation also covers lower poisoning rates. On GTSRB, the poisoning rate is 0.3% or lower depending on the target class, and TIDE still achieves the lowest ASR among all defenses for each of the five attacks (Tables 1–2). In these runs, the probe set typically contained only a small number of poisoned samples (about 1 to 3). On the larger Tiny ImageNet dataset, TIDE also reduces the ASR of Narcissus from 0.9899 to 0.2917 (Appendix H, Table 15), with the probe set typically containing about 5 poisoned samples.

Under the same approximation, at a poisoning rate of 0.05%, a probe size of at least 9,209 keeps the probability of containing no poison below 1%. Table 8 shows that more repetitions also improve TPR and reduce FPR. The cost grows linearly with the probe size (Table 7).

Effect of a lower poisoning rate. To study performance at a lower poisoning rate, we ran Narcissus on CIFAR-10 with 0.1% poisoning (50 poisoned samples) and report results averaged over three runs. In this experiment, we increased the probe size from 1,000 to 5,000 to obtain 5 poisoned probe samples in expectation. This probe-size choice uses the experimental poisoning rate, but the rate is not supplied to the peeling or cutoff-selection procedures.

Table R2 (Table 11 in Appendix E of the revised paper). Narcissus on CIFAR-10 with a 0.1% poisoning rate (50 poisoned samples) and a probe set of 5,000 samples, averaged over three runs.

UndefendedTIDEActivation ClusteringSPECTREFLARE
Poisons detected-28.70.00.00.0
Samples removed-1,867.00.00.00.0
Clean Acc.0.94710.94560.94610.94230.9387
ASR1.00000.64521.00001.00001.0000

TIDE is the only defense that detects any poisoned samples at this rate. It finds 57% of the poisoned samples on average, which reduces the ASR from 1.0000 to 0.6452 with almost no loss in clean accuracy. Other defenses leave the ASR at 1.0000. Although detection becomes more difficult as the poisoning rate decreases from 0.5% to 0.1%, TIDE still partially mitigates the attack in this challenging setting.

We have summarized the key evidence and its motivation from Appendix A into the main paper as a new Section 5.5 (Table 3). The section now summarizes the mean pairwise influence across seven settings (five attacks, two datasets), including a comparison with clean pairs restricted to the target class, and explains why this evidence motivates detection at the subgraph level. The full statistics (medians, tails, and cross-type pairs) remain in Appendix A (Table 4).

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

This paper proposes a data-cleaning defense against backdoor poisoning attacks. The main idea is that poisoned samples may be difficult to identify individually but tend to have stronger mutual influence during training. It estimates pairwise influence using a trajectory-based approximation, constructs a weighted bipartite graph, and applies a greedy peeling algorithm to identify suspicious samples. These samples are then removed before training the final model. Experiments on three datasets show that the proposed method can reduce attack success rates while maintaining relatively high clean accuracy.

Critical strengths
  1. Detecting poisoned samples without requiring a trusted clean reference dataset is a practically relevant setting.
  2. Identifying poisoned samples based on their mutual influence rather than individual feature anomalies is interesting. The graph-based formulation provides a different perspective on backdoor detection.
  3. The paper evaluates both detection performance and post-defense model performance, providing a useful assessment of the trade-off between attack mitigation and clean accuracy.
Critical weaknesses
  1. TIDE randomly partitions the dataset into source and probe sets. Since the poisoning rate is typically very low, each partition may contain only a small number of poisoned samples. How does this affect the reliability of influence estimation and detection?
  2. The method appears to work best when poisoned samples share a common trigger. If different poisoned samples contain different triggers, their mutual influence may be weaker. How would TIDE perform in this setting?
  3. The method assumes that poisoned samples have stronger mutual influence than clean samples. However, this assumption may depend on the trigger design, target class, and poisoning strategy. More analysis is needed to show when this assumption holds and when it may fail.
  4. The influence estimation relies on last-layer representations to approximate each sample's contribution to model updates. This may not fully capture interactions across different network layers. How reliable is this approximation, and how do estimation errors affect detection performance?
  5. TIDE requires additional training, repeated data partitions, and pairwise influence estimation. Although runtime results are provided, its scalability to larger datasets and more complex models remains unclear. A more detailed computational analysis would be helpful.
Questions for authors
  1. Since the poisoning rate is very low, random source-probe partitioning may leave only a few poisoned samples in the probe set. How sensitive is TIDE to the number and distribution of poisoned samples across the two sets?
  2. How would TIDE perform when poisoned samples contain different or sample-specific triggers rather than sharing a common trigger?
  3. The method assumes that poisoned samples have stronger mutual influence than clean samples. Under what conditions does this assumption hold? Could clean samples with similar features also form dense high-influence subgraphs?
  4. How accurate is the last-layer influence approximation? How would using influence information from multiple layers affect detection performance?
  5. How does TIDE scale to larger datasets and more complex architectures, given the additional training and pairwise influence estimation costs?
5 comments

Thank you for your feedback. The probe set is drawn uniformly at random, so its poison count follows a hypergeometric distribution, which we approximate here using a binomial distribution. With poisoning rate ρ\rho and probe size PP, the expected count is ρP\rho P.

In our main experiments (ρ=0.5\rho=0.5%, P=1,000P=1{,}000), a probe contains 5 poisons in expectation. Under the binomial approximation, the probability that it contains none is (1−ρ)P=0.9951000≈0.67(1-\rho)^P = 0.995^{1000} \approx 0.67% (<1<1%). The probability of at most one poison is about 4%, so a probe contains two or more poisons with approximately 96% probability. Moreover, TIDE aggregates over RR disjoint probes, so detection does not rely on any single partition. Under the same approximation, the probe size needed to keep Pr⁡\Pr(no poison) below 1% is 919 for ρ=0.5\rho=0.5%, 4,603 for ρ=0.1\rho=0.1%, and 9,209 for ρ=0.05\rho=0.05%.

To examine sensitivity to the distribution of poisons across the two sets, we compared random partitions with partitions in which all poisons are forced into the source set, so that the probe contains no poison (same setting as our main experiments, averaged over three runs). With a random probe, TIDE achieves TPR 0.7260 and FPR 0.2597, and with a probe containing no poison, TPR 0.4045 and FPR 0.1744. Detection is therefore weaker in this worst case, which occurs with probability below 1% in our main setting, but it does not collapse.

To examine sensitivity to the number of poisons, we also ran Narcissus on CIFAR-10 with a 0.1% poisoning rate (50 poisoned samples) and a probe of 5,000 samples, averaged over three runs. TIDE detects 28.7 of the 50 poisoned samples on average and reduces the ASR from 1.0000 to 0.6452, whereas Activation Clustering, SPECTRE, and FLARE detect none and leave the ASR at 1.0000. We report these analyses in the new Appendix E (Table 11).

TIDE does not require a single shared trigger. It relies on poisoned samples influencing each other more strongly than clean samples influence each other, and it detects this at the subgraph level rather than from individual pairs. Our evaluation already includes COMBAT, whose trigger is generated per input rather than shared. TIDE still detects it (e.g., for CIFAR-10, TPR 0.6352 at FPR 0.0729, reducing ASR from 0.8402 to 0.1789), and Section 5.5 (Table 3) confirms that poison-to-poison influence remains the strongest pair type there, although with a narrower relative margin than for Narcissus, which has a shared trigger.

To examine how this pattern varies across poisoning strategies and datasets, we computed pairwise influence statistics for seven settings (five attacks, two datasets), including clean pairs restricted to the target class (Table R1; Section 5.5 and Appendix A, Table 4). Poison-to-poison pairs have the highest mean influence in every evaluated setting, exceeding clean-to-clean pairs both within the target class and across all classes. The observed pattern therefore extends across the evaluated attacks rather than being confined to a single trigger design.

We also compare clean-to-clean pairs within the target class with clean-to-clean pairs across all classes to examine whether target-class samples are associated with higher mean influence. Target-class clean pairs have higher mean influence than clean pairs across all classes in all seven settings, but their means remain below those of poison-to-poison pairs. The separation varies across settings. For example, the poison-to-poison and target-class clean means are 0.7136 versus 0.3655 for CIFAR-10 Wicked-BadNet, compared with 8.5550 versus 7.5703 for Tiny ImageNet Wicked-BadNet. The ordering is not consistent across all statistics (e.g., the target-class clean median exceeds the poison-to-poison median for CIFAR-10 Narcissus, and target-class clean pairs have a heavier tail for COMBAT), indicating that the mean advantage does not imply uniformly stronger poison-to-poison interactions. This is why TIDE detects poisoned samples at the subgraph level rather than from individual pairwise values.

Table R1. Pairwise influence statistics (a subset of Table 4 in Appendix A). Clean denotes all clean samples, and Clean_t denotes clean samples from the target class.

PairMeanMedianStd. dev.95th pct.
CIFAR-10, Wicked-BadNet
Poison →\to Poison0.71360.12991.36623.4980
Clean →\to Clean0.25530.06360.53761.2090
Clean_t →\to Clean_t0.36550.08930.78051.7207
CIFAR-10, Wicked-Blended
Poison →\to Poison0.68810.13051.34033.3828
Clean →\to Clean0.36860.08970.79181.7412
Clean_t →\to Clean_t0.47170.20430.75421.8516
CIFAR-10, Wicked-SIG
Poison →\to Poison0.67670.09761.39353.4316
Clean →\to Clean0.23240.10470.42030.8604
Clean_t →\to Clean_t0.24440.05680.52891.1729
Tiny ImageNet, Narcissus
Poison →\to Poison9.35567.95517.452023.2828
Clean →\to Clean7.37625.43366.788620.6094
Clean_t →\to Clean_t7.55486.15626.253219.4375
Tiny ImageNet, Wicked-BadNet
Poison →\to Poison8.55505.85169.352326.3852
Clean →\to Clean7.54465.47857.441322.1344
Clean_t →\to Clean_t7.57036.22666.170119.2656

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.