Pith.

Cryptography and Security · arXiv:2611.09348 v1 · submitted 6 Nov 2026 · 20 references

When Feedback Becomes a Backdoor: Tracing Backdoor Skill Formation in Text-Space Skill Optimization

Louise Nayeong Koh, Yaechan Lee, Myeongjong Lee, Doyeon Kim, Juyoung Park, Seunggyu Kim, Inseo Kim, Chaerin Jung

show 1 more author

Hyogon Ryu

The paper says

“Reference answers are a supervision channel in text-space Skill optimization, which produces reusable text artifacts called Skills that guide future agent behavior.”

more of the abstract

Poisoning those answers can induce a persistent trigger-conditioned rule without giving the attacker access to the model, optimizer, or Skill file. We study this reference-feedback poisoning in SkillOpt by tracing an attacker-supplied trigger–payload mapping from reflection to held-out activation. In SearchQA, where the payload replaces a short answer, the natural attack retains an exact rule in all three seeds; each final Skill emits the payload on 200/200 triggered questions and 0/200 clean questions. In seed 42, removing the rule eliminates activation, and an in-run comparison shows similar clean accuracy to benign SkillOpt (169/200 versus 171/200). In SpreadsheetBench, where the payload requires an additional workbook edit, no natural run retains an exact rule across three seeds. Diagnostic interventions locate two successive bottlenecks: the Analyst abstracts away required literals, and, once literal preservation is enabled, edit-budget ranking can remove exact edits before candidate construction. Poisoned references can therefore create persistent backdoors, but exposure alone does not ensure that the induced rule is formed, selected, and retained.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 9 Nov 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. A genuinely new attack channel, studied with unusual care; the three-seed persistence result holds for one aligned construction, and the arbitrary-payload case is a disclosed single-seed existence proof. This line is wrong

1 major objection · 6 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 20 references say.

The paper says, at Section 5.4 and Table 12; abstract

The review did not quote this passage.

A classifier read this sentence of the paper as acknowledging the concern (probability 0.64):

“Limitations.Natural end-to-end success covers one SearchQA mapping across three seeds.”

Pith says: Objection 1 of 1 · major

The abstract's concluding claim that "poisoned references can therefore create persistent backdoors" is broader than the empirical basis. The three-seed natural end-to-end result uses only the aligned mapping [ZEPHYR-17] → ZEPHYR VERIFIED (Table 5). The lexically unrelated payload ORCHID-729 completes the full chain in one seed and only under the stronger condition with 10% poisoned validation references (Section 5.4), whereas under the basic condition the opaque-tag construction is payload-free in the endpoint Skills across seeds (Table 12).

more of objection 1

Because Section 5.4 itself uses the opaque payload to answer the "backdoor or format rule?" question, this single trajectory is load-bearing for the headline claim. The Limitations paragraph acknowledges the scope, but the abstract and the contributions list do not carry the qualification. I recommend either running the opaque construction under the stronger condition across at least three seeds, or explicitly stating in the abstract and conclusion that the three-seed persistence result concerns the aligned mapping and that an arbitrary opaque payload has been demonstrated only in a single trajectory.

6 minor objections

The paper says, at Abstract

Pith says: minor objection 1 of 6

The abstract contains a spacing typo: "calledSkills" should be "called Skills."

The paper says, at Section 4.1

Pith says: minor objection 2 of 6

The citation formatting is inconsistent: "SearchQA[ 11]" and "SpreadsheetBench[ 12]" have a space before the bracket, unlike the reference citations elsewhere.

show 4 more

The paper says, at Section 5.2.3 and Appendix G

Pith says: minor objection 3 of 6

The validation threshold q_best = 0.175 is used implicitly in the main text but defined only in Appendix G; it should be defined at first use in Section 5.2.3 so the gate-pass discussion is self-contained.

The paper says, at Section 5.2.3 and Appendix G

Pith says: minor objection 4 of 6

The external-carrier conclusion is drawn from a 40-item positive-control set deliberately chosen because the parent answers inconsistently. The main text should state more prominently that this is a mechanism demonstration on a crafted set, not an estimate of how often clean edits carry backdoor edits in natural optimization runs.

The paper says, at References

Pith says: minor objection 5 of 6

Reference [15] has a typo: "V oyager" should be "Voyager."

The paper says, at Appendix J

Pith says: minor objection 6 of 6

The paper states that code and run artifacts are not publicly released. Given that the target model is closed and the provider does not expose checkpoint details, independent replication is currently impossible; I would encourage the authors to release at least a minimal evaluation harness, frozen Skills, and redacted traces, or to explain why this is not possible.

What would settle it

Re-run the Section 5.4 ORCHID-729 configuration (100 training examples, 10 poisoned references, 10 poisoned validation references) on seeds 43 and 44 with identical SkillOpt settings, and repeat the seed-42 run on an open-weight model (e.g., Llama-4 or Qwen-3) using the same optimizer prompts and edit budget. Criterion: if ORCHID-729 is retained in the endpoint Skill and yields 200/200 triggered and 0/200 clean activation in at least two of the three seeds, or on the open-weight model, the arbitrary-payload persistence claim is robust; if seeds 43/44 and the open model all fail to retain the opaque rule, the headline should be scoped to the aligned mapping under the basic condition plus a one-seed existence result for opaque payloads.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 9 Nov 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic agrees 1 min 19 s · 17,311 tokens read, 12,270 written
  3. Editor wrote the verdict above 47 s · 18,188 tokens read, 7,399 written
  4. Referee major revision: 1 major objection, 6 minor 58 s · 17,588 tokens read, 9,509 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

The short version: poisoning reference answers, with no access to the model, optimizer, or Skill file, can plant a persistent, selectively triggered backdoor in SkillOpt. The paper traces a trigger-payload rule through every promotion stage and shows it surviving end-to-end in three seeds on SearchQA: 200/200 triggered, 0/200 clean, with a rule-deletion control. I buy the central mechanism, and the paper is unusually careful about what counts as evidence.

What's new is the channel. AgentPoison and MINJA poison memory or retrieval; OEP, SkillJack, PoisonedEvolution and TBA poison trajectories or queries; Skill-Inject assumes control of the Skill file; Rando and Tramèr poison RLHF preference labels. Scoring references sit outside all of those, and the paper separates the training-reference role (formation) from the validation-reference role (selection). The stage-resolved promotion analysis is a contribution in its own right, and the carrier effect — a validation-neutral backdoor edit riding into the best Skill on the gains of an unrelated clean edit — is a genuine mechanistic finding.

The SpreadsheetBench negative result is handled well. Instead of burying it, they localize two gates (Analyst abstraction, then edit-budget clipping) with diagnostics, and never count diagnostics toward natural success. They also caught and reported a false activation caused by non-isolated endpoint evaluation. That's the right instinct, and it makes the positive results easier to trust.

Soft spots, in order of size. First, the persistence claim leans on one construction. The aligned mapping [ZEPHYR-17] → ZEPHYR VERIFIED was chosen after a one-seed probe showed it was the only one of four constructions that survived the Analyst's abstraction, and it is semantically aligned — which is why the opaque payload ORCHID-729 matters. That payload completes the chain in one seed and only under the stronger condition with 10% poisoned validation references; under the basic condition its endpoints are payload-free. So the arbitrary-payload case is an existence proof, and the authors say as much. The abstract's 'can create persistent backdoors' is fair; the evidence just doesn't support more generality.

Second, everything runs on one closed model, GPT-5.6 Luna, with no checkpoint or decoding details and no public artifacts. For a claim about a class of optimizers, that's the binding constraint on transfer. The authors disclose it; it's still the main reason to treat the headline as provisional.

Third, small: the clean-utility comparison (169/200 vs 171/200) is seed 42 only, and TBA only runs on SpreadsheetBench. Both disclosed, both acceptable scoping.

Who this is for: people working on self-evolving agent security, and anyone designing poisoning evaluations — the stage-resolved protocol is worth copying. It deserves serious refereeing. The asks I'd make: release artifacts, add a second model or a second arbitrary mapping, and run the opaque payload across seeds under the stronger condition or say plainly it's a single-seed existence result.

The strongest objection

The central claim is that reference-feedback poisoning can create persistent backdoors, and the paper's own standard for the main result is three-seed reproducibility. That standard is met only for the aligned mapping [ZEPHYR-17] -> ZEPHYR VERIFIED under the basic condition (Table 5; [b3], [b105]). The only lexically unrelated payload, ORCHID-729, completes the chain in one seed and only under the stronger condition with 10% poisoned validation references (Section 5.4; [b178]); under the basic condition its endpoints are payload-free across seeds (Table 12). Since the aligned mapping was selected after a one-seed probe and all runs use one closed model without checkpoint or decoding details ([b77], [b183]), the persistence result could be specific to this construction and to this model's abstraction and selection behavior. The authors disclose this scope ([b176], [b183]), so the issue is not internal inconsistency; it is that the headline 'poisoned references can create persistent backdoors' currently leans on a single trajectory for the arbitrary-payload case.

The reader's conditional verdict is built on the same generality concern, and my read agrees. Within the paper, the stage-resolved evidence is coherent: the SpreadsheetBench natural negative is converted into a mechanism analysis rather than hidden, the carrier and deletion controls support the mechanism, and diagnostics are correctly excluded from natural counts. I found no internal inconsistency that would justify REJECT. The residual risk is external validity of the strongest reading of 'persistent backdoors': the only arbitrary-payload success is a single seed under validation poisoning, and the three-seed natural result uses a construction chosen after a one-seed probe on one closed model. The proposed check directly tests whether that reading survives replication; until then CONDITIONAL is the right verdict, unchanged by this pass.

The referee's report

If the result holds, the paper identifies a meaningful and previously understudied attack surface: reference answers, not just trajectories, memory stores, or Skill files, can be the carrier of a persistent trigger-conditioned rule. The study is methodologically careful in several respects: it separates natural end-to-end evidence from mechanism diagnostics and never counts diagnostics toward attack success; it provides direct behavioral measurement against clean baselines; and it includes a natural negative (SpreadsheetBench) that disciplines the positive SearchQA result. The stage-resolved evaluation is a useful contribution in itself, and the candidate-level carrier effect is an interesting mechanistic finding. The main weakness is external scope: all runs use one closed model, the three-seed natural success uses a single semantically aligned mapping, and the only opaque-payload full chain is a single trajectory under the stronger condition. These constraints are acknowledged in the Limitations paragraph, but the abstract and conclusions state the backdoor claim more broadly than the evidence supports.

Pith's reading of the paper (a machine's summary, not the authors')

Self-improving agents refine their execution experience into persistent text instructions called Skills, and a text-space Skill optimizer supervises that refinement by scoring executions against reference answers. This paper asks whether an attacker who supplies only tasks and reference answers — with no access to the model, optimizer, or Skill file — can plant a trigger-conditioned rule that survives optimization and fires later on inputs carrying the trigger. The answer is conditionally yes: in a short-answer QA setting the natural attack retained an exact trigger-to-payload rule in all three seeds, with each final Skill emitting the payload on 200/200 held-out triggered questions and 0/200 clean questions at roughly unchanged clean accuracy (169/200 versus 171/200 in the matched seed). In a spreadsheet-editing setting the same attack never persisted, and diagnostics located the two gates that stopped it: the Analyst's tendency to abstract away exact literals, and edit-budget ranking that removes exact edits before a candidate is built. The paper's overall claim is that poisoned references can create persistent backdoors, but exposure to the poison alone does not ensure that the induced rule is formed, selected, and retained.

What carries the argument. The load-bearing object is the promotion-stage chain, an eight-transition path a poisoned rule must complete to become a persistent backdoor: recognition in reflection, exact proposal as an edit, survival through merging, survival through ranking, entry into a candidate Skill, beating the incumbent on validation, presence in the endpoint Skill, and activation on held-out inputs. The paper reads each stage from immutable run artifacts instead of from the final Skill alone, which is what allows it to separate the claim "the poison was seen" from the claim "the rule persisted." A second mechanism, the carrier effect, explains how an exact but validation-neutral backdoor edit can cross the validation gate: because SkillOpt accepts or rejects a candidate atomically, an unrelated clean edit that raises the validation score can carry the backdoor edit into the best Skill with it — the score change attributable to the backdoor itself is about 0.0050, versus about 0.2075 for the carrier edit.

Core claim. The central discovery is a complete natural backdoor chain through text-space Skill optimization achieved by an attacker who controls nothing but tasks and reference answers. When a poisoned training reference replaces the answer to a trigger-bearing question, the scored executions change which behaviors are labeled as failures, reflection proposes an edit encoding the rule $[\text{ZEPHYR-17}]\to\text{ZEPHYR VERIFIED}$, and that exact rule can survive merging, ranking, candidate construction, validation, and retention to sit in the endpoint Skill. The paper reports this full chain in all three SearchQA seeds: each final Skill emits the payload on 200/200 held-out triggered questions and on 0/200 clean questions, deleting the retained rule eliminates the behavior, and the poisoned Skill's clean accuracy is comparable to a benign run. It reports the opposite face in SpreadsheetBench, where the same attack retains no exact rule across three seeds and stage-resolved diagnostics isolate two successive bottlenecks, literal abstraction by the Analyst and edit-budget clipping. The general claim is that persistence is a multi-stage outcome: poisoned references can complete it, but nothing about exposure guarantees the later stages.

The premise it rests on. The persistence result is carried by a single trigger-payload construction — the aligned mapping $[\text{ZEPHYR-17}]\to\text{ZEPHYR VERIFIED}$ — evaluated on one closed model whose checkpoint and internal abstraction behavior are not inspectable; if that mapping's survival through the Analyst is a quirk of this model or wording, the claim that reference-feedback poisoning produces persistent backdoors would not transfer to other models or payloads. The paper states this scope itself in its Limitations paragraph: natural end-to-end success covers one SearchQA mapping across three seeds, and results come from one optimizer and one closed model.

If the paper is right (Pith's inferences, not the authors')

A dataset contributor with no access to the model, optimizer, prompts, or Skill file can implant a selectively activated rule into a self-improving agent's persistent Skill; in SearchQA this completed naturally in all three seeds, and the support grid shows success even at 50 training examples with only 5 poisoned references.

Exposure to poisoned references is not persistence: recognition, proposal, merging, ranking, candidate entry, validation, retention, and activation are independent gates, so attack evaluations need stage-resolved reporting rather than a single success number.

A persistent backdoor can leave observed clean utility nearly untouched, because the carrier effect lets a validation-neutral backdoor edit ride into the best Skill on an unrelated clean edit's validation gains (169/200 versus 171/200 clean accuracy in the matched seed).

The same poisoning recipe does not transfer across task settings: in the workbook-editing environment the natural chain stalled at formation or ranking in all three seeds, localizing defense-relevant gates in the Analyst's literal abstraction and in edit-budget ranking.

Retained rules varied by seed from near-literal to family-wide trigger scope (0/200 to 200/200 activation on unseen in-family triggers, 0/200 on clean and out-of-family inputs), so the breadth of an implanted backdoor is not fully under the attacker's control.

Editorial inference: the seed-dependent trigger scope means a defender could audit final Skills for trigger families notably broader than the training distribution justifies; the paper's own data show a family rule firing on 200/200 unseen in-family triggers, a signature a purely literal poison would not produce.

Editorial inference: the carrier effect directly motivates leave-one-edit-out validation — rescoring each candidate Skill with one edit removed — which the paper lists as a promising defense but does not evaluate; its detection rate and clean-utility cost are a natural, as-yet-unmeasured extension.

Editorial inference: because every run uses one closed model whose checkpoint and accelerator details are unavailable, an open-weight replication of the aligned-mapping attack is the cleanest transfer test; a single seed that fails to retain the exact payload would narrow the result from a property of reference-feedback poisoning to a property of one model's abstraction behavior.

Editorial inference: the paper's own control — apparent SpreadsheetBench activations that vanished after repository and prior-state access were removed — implies that hosted agent evaluations should require isolated workspaces and execution traces before reporting backdoor activation, since prior-state leakage can fake the phenomenon being measured.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Reference answers are a supervision channel in text-space Skill optimization, which produces reusable text artifacts called Skills that guide future agent behavior. Poisoning those answers can induce a persistent trigger-conditioned rule without giving the attacker access to the model, optimizer, or Skill file. We study this reference-feedback poisoning in SkillOpt by tracing an attacker-supplied trigger–payload mapping from reflection to held-out activation. In SearchQA, where the payload replaces a short answer, the natural attack retains an exact rule in all three seeds; each final Skill emits the payload on 200/200 triggered questions and 0/200 clean questions. In seed 42, removing the rule eliminates activation, and an in-run comparison shows similar clean accuracy to benign SkillOpt (169/200 versus 171/200). In SpreadsheetBench, where the payload requires an additional workbook edit, no natural run retains an exact rule across three seeds. Diagnostic interventions locate two successive bottlenecks: the Analyst abstracts away required literals, and, once literal preservation is enabled, edit-budget ranking can remove exact edits before candidate construction. Poisoned references can therefore create persistent backdoors, but exposure alone does not ensure that the induced rule is formed, selected, and retained.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-11-09

The authors' main claim, as the classifier picked it out 1.00

“Poisoned references can therefore create persistent backdoors, but exposure alone does not ensure that the induced rule is formed, selected, and retained.”

How the main claim is stated: asserted 0.69

What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.87 · human participants 0.03 · controlled comparison 0.90 · randomized assignment 0.03 · states quantitative result 0.98 · compares with prior results 0.71 · mentions code or data availability 0.03

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Reference answers are a supervision channel in text-space Skill optimization, which produces reusable text artifacts called Skills that guide future agent behavior. background 1.00 0.09 0.03 asserted 1.00
Poisoning those answers can induce a persistent trigger-conditioned rule without giving the attacker access to the model, optimizer, or Skill file. background 0.81 0.17 0.18 asserted 0.75
We study this reference-feedback poisoning in SkillOpt by tracing an attacker-supplied trigger--payload mapping from reflection to held-out activation. contribution 0.89 0.92 0.07 asserted 0.99
In SearchQA, where the payload replaces a short answer, the natural attack retains an exact rule in all three seeds; each final Skill emits the payload on 200/200 triggered questions and 0/200 clean questions. result 1.00 0.93 0.08 asserted 1.00
In seed 42, removing the rule eliminates activation, and an in-run comparison shows similar clean accuracy to benign SkillOpt (169/200 versus 171/200). result 1.00 0.93 0.05 asserted 1.00
In SpreadsheetBench, where the payload requires an additional workbook edit, no natural run retains an exact rule across three seeds. result 1.00 0.91 0.04 asserted 1.00
Diagnostic interventions locate two successive bottlenecks: the Analyst abstracts away required literals, and, once literal preservation is enabled, edit-budget ranking can remove exact edits before candidate construction. result 0.82 0.92 0.09 asserted 1.00
Poisoned references can therefore create persistent backdoors, but exposure alone does not ensure that the induced rule is formed, selected, and retained. interpretation 0.97 0.81 0.81 asserted 0.69

What kind of paper the classifier reports this is: empirical study 1.00 · evidence from observation or experiment 0.97

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-09: 1 request, 6,354 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (2)
Figure 1 of the paper
Figure 1. Victim optimization loop. Execution is scored against references; reflection proposes edits;
Figure 2 of the paper
Figure 2. Threat model for reference-feedback poisoning. The attacker supplies a subset of training

Figures are the authors', reproduced from the paper.

References (20)
  1. [1] Y . Yang et al. SkillOpt: Executive Strategy for Self-Evolving Agent Skills.arXiv preprint arXiv:2605.23904, 2026 on PithPith's reviewarXiv
  2. [2] Z. Ying et al. SkillJack: Persistent Skill Backdoors in Self-Evolving Agents.arXiv preprint arXiv:2608.03509, 2026 on PithPith's reviewarXiv
  3. [3] Y . Luo et al. Query-Only Backdoor Attacks on Self-Evolving Skills via Trajectory Poisoning. arXiv preprint arXiv:2608.08303, 2026 on PithPith's reviewarXiv
  4. [4] J. Chen et al. When Experience Becomes Instruction: Trajectory Poisoning in Self-Evolving Agent Skill Systems.arXiv preprint arXiv:2608.05563, 2026 on PithPith's reviewarXiv
  5. [5] K. Wang et al. OEP: Poisoning Self-Evolving LLM Agents via Locally Correct but Non- Transferable Experiences.arXiv preprint arXiv:2605.18930, 2026 on PithPith's reviewarXiv
  6. [6] D. Schmotz et al. Skill-Inject: Measuring Agent Vulnerability to Skill File Attacks.arXiv preprint arXiv:2602.20156, 2026 on PithPith's reviewarXiv
  7. [7] Z. Chen et al. AgentPoison: Red-Teaming LLM Agents via Poisoning Memory or Knowledge Bases.Advances in Neural Information Processing Systems (NeurIPS), 2024 on Pith
  8. [8] S. Dong et al. Memory Injection Attacks on LLM Agents via Query-Only Interaction (MINJA). Advances in Neural Information Processing Systems (NeurIPS), 2025 on Pith
  9. [9] S. Shao, Q. Ren, D. Liu, et al. Your Agent May Misevolve: Emergent Risks in Self-evolving LLM Agents.International Conference on Learning Representations (ICLR), 2026 on Pith
  10. [10] J. Rando and F. Tramèr. Universal Jailbreak Backdoors from Poisoned Human Feedback. International Conference on Learning Representations (ICLR), 2024 on Pith
  11. [11] M. Dunn, L. Sagun, M. Higgins, V . U. Guney, V . Cirik, and K. Cho. SearchQA: A New Q&A Dataset Augmented with Context from a Search Engine.arXiv preprint arXiv:1704.05179, 2017 on PithPith's reviewarXiv
  12. [12] Z. Ma et al. SpreadsheetBench: Towards Challenging Real World Spreadsheet Manipulation. Advances in Neural Information Processing Systems (NeurIPS), 2024 on Pith
  13. [13] N. Shinn et al. Reflexion: Language Agents with Verbal Reinforcement Learning.Advances in Neural Information Processing Systems (NeurIPS), 2023 on Pith
  14. [14] A. Zhao et al. ExpeL: LLM Agents Are Experiential Learners.AAAI Conference on Artificial Intelligence, 2024. 15 on Pith
  15. [15] G. Wang et al. V oyager: An Open-Ended Embodied Agent with Large Language Models. Transactions on Machine Learning Research, 2024 on Pith
  16. [16] M. Yuksekgonul et al. Optimizing generative AI by backpropagating language model feedback. Nature, 639:609–616, 2025 on Pith
  17. [17] A. Wan, E. Wallace, S. Shen, and D. Klein. Poisoning Language Models During Instruction Tuning.International Conference on Machine Learning (ICML), 2023 on Pith
  18. [18] A. Schwarzschild et al. Just How Toxic is Data Poisoning? A Unified Benchmark for Backdoor and Data Poisoning Attacks.International Conference on Machine Learning (ICML), 2021 on Pith
  19. [19] T. Gu, K. Liu, B. Dolan-Gavitt, and S. Garg. BadNets: Evaluating Backdooring Attacks on Deep Neural Networks.IEEE Access, 7:47230–47244, 2019 on Pith
  20. [20] Y . Zeng, W. Park, Z. M. Mao, and R. Jia. Rethinking the Backdoor Attacks’ Triggers: A Frequency Perspective.IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 16 on Pith

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment