acceptodds
Under review as a conference paper at ICLR 2027

From Triggers to Antigens: An Immune-Inspired Framework for Continual Backdoor Defense

Abstract

Deep neural networks are increasingly deployed in dynamic environments where models must continuously learn from evolving data streams, yet many existing backdoor defenses remain fundamentally limited by their reliance on predefined trigger patterns. Such a trigger-centric view fails when attacks become distributed, representation-dependent, or semantically consistent with their target labels, leaving the underlying malicious behavior invisible to conventional detectors. In this work, we reveal a different perspective: backdoors are not characterized by their observable triggers, but by recurring non-self behaviors that persist after normal semantic responses are removed. Through extensive empirical analysis, we discover that samples carrying a recurring backdoor factor across different semantic contents can share a class-conditioned residual, analogous to a pathogen-derived antigen exposed after subtracting healthy self responses. This observation motivates a new formulation of continual backdoor defense as an antigen-to-antibody learning process rather than a trigger identification problem. Building upon this insight, we propose Immune-inspired Semantic Network (ISN), a continual defense framework that learns transferable immune memory against emerging backdoor behaviors. ISN establishes a stable semantic self-reference, extracts attack-invariant antigens through complementary receptors, and transforms validated antibodies into adaptive immune responses including correction, neutralization, and vaccination. Unlike many existing approaches that memorize individual attacks, ISN accumulates reusable immune memory capable of recognizing recurrent threats across tasks and attack families. Empirical results demonstrate that our framework achieves the state-of-the-art performance.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
machine learning, continual learning, continual defense

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_h4Y2OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes ISN, an immune-inspired framework for continual backdoor defense. The key idea is to extract shared class-conditioned residuals from semantically conflicting poisoned samples as “antigens,” learn corresponding “antibodies,” and reuse them to detect and mitigate recurrent backdoor behaviors across tasks. The framework combines a frozen semantic encoder, multiple receptor types, training-time correction/neutralization/vaccination, and immune memory. Experiments on continual image classification benchmarks show strong reductions in attack success rate while largely preserving clean accuracy.

Critical strengths
  • The paper presents a different perspective on continual backdoor defense by modeling shared residual patterns rather than focusing only on explicit trigger appearance.

  • The proposed framework covers several stages of the defense pipeline, including semantic screening, antibody construction, mitigation, and cross-task memory.

  • The experiments suggest that the method can reduce ASR while maintaining competitive clean accuracy on the evaluated benchmarks.

Critical weaknesses
  • My main concern is that ISN needs high-confidence semantic conflicts as antigen seeds before it can fit a new antibody. This creates a tension with one of the paper's main motivations, namely detecting semantically consistent or clean-label backdoors. If a new attack only produces label-consistent poisoned samples and therefore provides few or no semantic-conflict seeds, ISN cannot establish a new antibody unless a compatible mature memory already exists.

  • No adaptive attack against ISN is considered in this paper.

  • The experiments mainly show transfer across different source classes under the same attack configuration. For example, the antibody learned from poisoned bird/horse samples is tested on poisoned target-class samples carrying the same BadNets or WaNet trigger. Therefore, it is still unclear whether the learned “antigen” captures a trigger-independent malicious behavior, or simply a class-conditioned signature of that specific trigger. A stronger evaluation would vary the trigger pattern, location, or attack realization while keeping the target behavior unchanged, and test whether the same antibody still transfers.

  • The claim that immune memory transfers across attack families is not directly demonstrated. Different attacks are mainly handled by different receptor pathways, which is weaker than showing that an antibody learned from one attack can transfer to another.

  • The method relies heavily on the frozen semantic encoder to identify conflicts, define clean references, and assign pseudo-labels. Its robustness when the encoder itself makes semantic mistakes is not sufficiently studied.

  • The setting is limited to class-disjoint Task-IL with known task identity at inference. It does not cover harder settings such as Class-IL or unknown task identity.

  • The attack coverage is relatively narrow. The main experiments focus on BadNets, WaNet, and LTB, with fixed poisoning settings; more diverse poison rates, targets, dynamic triggers, and multiple simultaneous backdoors are missing.

  • The paper seems to rely quite heavily on manual \vspace adjustments to fit the content, which leads to several noticeable spacing issues

Questions for authors

see weaknesses

7 comments
AuthorsOpenReview ↗

Response to Reviewer h4Y2 #1

Dear Reviewer h4Y2,

Thank you for the detailed review and for recognizing the shared-residual perspective and the observed defense performance. We have added controls that directly exercise the acquisition bottleneck, evaluate a defense-aware search with an active antibody, and replay its frozen candidate on newly fitted victims. We also use the submitted mechanism experiments and new reference interventions to clarify what is transferred and how semantic errors affect training. The added experiments use CIFAR-10 and seed 42. We identify unconditional ASR and the paper's clean-conditioned ASR separately throughout.

AuthorsOpenReview ↗

Response to Reviewer h4Y2 #2

W1. Acquiring an antigen when the semantic discovery gate supplies no usable hypothesis

The revised experiment creates the failure mode you identify under an effective attack and shows that a complementary acquisition route restores protection. We separate the availability of semantic evidence from whether an attacker has changed the ground-truth label. The former is the operational input to Eq. (4).

The submitted mechanism first transfers evidence across samples within an attack. Sec. 3.3, p. 4, describes that process:

Immune perception operationalizes this evidence transfer in three steps: it exposes cross-class seeds, fits their recurring non-self response through complementary receptors, and uses validated antibodies to screen same-class data.

Thus, label-consistent poisoned samples are a population to be recognized after a detector has been acquired. They need not reveal themselves through label disagreement. Sec. 3.3, p. 6, states:

Semantic consistency provides the retained label, and antibody evidence determines intervention.

In the additional first-task experiment with reliable seeds, poisoned Vanilla has 92.48% conditional ASR, while ISN has 0/1,000 at 99.95% CA. ISN starts with empty memory, obtains 800 poisoned examples among 805 conflict seeds, and admits one local antibody. This establishes that historical memory is unnecessary when current evidence supports acquisition.

We then test natural failure of the current seed gate. The first task contains 9,680 training images and 968 label-flipping poisons with a fixed 3×3 white patch. With its fixed reference split, the attack target produces zero semantic-conflict seeds. The other class produces five clean conflicts, below the minimum support of six. Neither class provides an accepted hypothesis. The original ISN then acquires the backdoor, with 92.00% unconditional ASR and 91.98% conditional ASR.

To address this bottleneck, the complementary route proposes recurring local patterns from images and observed labels, checks recurrence on a separate training fold, and screens candidates against trusted clean references. This route obtains candidate evidence directly from the stream without consuming conflict seeds. The filtered stream then enters the ISN learner.

First-task conditionCA (%)Unconditional ASR (%)Conditional ASR (%)
Clean ISN99.900.100.00
Poisoned ISN99.8092.0091.98
Poisoned ISN with class-matched random removal99.8091.9091.88
Poisoned ISN with revised screening99.950.100.00

The screen and random-removal arms delete exactly 1,026 images under the same observed-class quotas, retain 8,654 images, and perform 2,710 main optimizer updates. Screening removes all 968 poisons and 58 clean examples. Random removal removes 164 poisons, leaving 804 in training. The 91.80-percentage-point unconditional ASR gap between the two arms therefore isolates selection quality at the fixed retained-data and update budget. Poison identities are used retrospectively to count removals.

The combined acquisition experiment uses 120 distinct trusted examples per class, including 60 additional examples beyond the three-split supplementary implementation. We report this budget with the result. The test uses label-flipping poisons and naturally absent usable target seeds. The acquisition rule itself uses images, observed labels, recurrence, and clean references. These inputs also define how the branch operates when the observed labels are correct. We will state the empirical composition and the discovery condition separately in the revised method and abstract.

We will incorporate the complementary acquisition rule into Sec. 3.3 and state its inputs, recurrence checks, clean calibration, and trusted-data budget. Appendix E.2 supplies the separate test of evidence transfer from an exposed subset to hidden samples carrying the learned residual.

AuthorsOpenReview ↗

Response to Reviewer h4Y2 #3

W2. Defense-aware search with an active antibody and newly fitted victims

We now test an attacker that jointly seeks target induction, semantic-conflict evasion, and affinity evasion against a surrogate containing an active target antibody. This directly exercises the detector state relevant to your concern.

The surrogate admits one local spatial antibody for the target. The attacker performs a single-start, 40-step SPSA search, with 121 optimization queries and two verification queries. Its objective combines target cross-entropy with penalties for semantic exposure and antibody affinity. Candidate selection uses training-side data, followed by frozen-candidate evaluation.

Across the 40 recorded updates, affinity penalties decrease while target cross-entropy increases. The initial joint loss is 6.62286, and the smallest loss among the recorded updates is 6.68086. The selection procedure therefore retains the initial white patch. This trajectory records the actual optimization trade-off. A successful evasion candidate must preserve target induction as well as lower detector response.

On an independent reporting holdout from the training data, the fixed surrogate detects all 20 triggered source-class examples in the union of its two signals. Semantic conflicts cover 18/20, and the antibody covers 5/20. These counts show how the complementary detection signals operate on the same attack cohort.

We then freeze the selected candidate and repeat the four-arm victim evaluation with complete separation of trusted references. Each victim's anchor, validation, and calibration sets have zero overlap with every trusted-reference role, optimization set, and selection set of the source surrogate. The victims share the architecture, pretrained backbone, underlying corpus, and training seed with the surrogate. New LoRA parameters and classification heads are fitted after the candidate is frozen. Each checkpoint receives one official first-task test evaluation after its training-side counts are reproduced.

Newly fitted victimTraining conditionCA (%)Unconditional ASR (%)Conditional ASR (%)
Vanilla LoRAClean99.950.00 (0/1,000)0.00 (0/1,000)
Vanilla LoRAPoisoned99.8092.00 (920/1,000)91.98 (917/997)
ISNClean99.900.00 (0/1,000)0.00 (0/999)
ISNPoisoned99.900.00 (0/1,000)0.00 (0/999)

Each victim uses 60 trusted examples per class (120 total), split into 20 anchors, 20 validation examples, and 20 calibration examples per class. It trains for ten epochs on 8,680 images, with 868 poisons delivered in each poisoned arm. Each uses SGD 0.01, batch size 32, gradient accumulation 1, and 2,720 main optimizer updates. ISN's auxiliary heating stage is recorded separately.

Poisoning raises the undefended victim's unconditional ASR by 92.00 points. Under the same selected trigger, ISN reduces 920/1,000 successful attacks to 0/1,000 and raises CA from 99.80% to 99.90%. Clean and poisoned ISN both achieve 99.90% CA and zero successful attacks. These four arms quantify attack learning and the defense effect with fully separated trusted references.

We will report the search budget, objective, active-antibody state, selected pixels, and reference relationships explicitly. The selected patch remains the initial patch after the specified search. The result records the outcome of that defense-aware search and its transfer to newly fitted victims.

Reviewer_f2QAOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper connects the antigen-to-antibody process in adaptive immunity with continual backdoor defense, and proposes ISN to address an asymmetry that trigger-centric detectors cannot handle. Poisoned samples whose labels conflict with their semantics are already exposed, whereas label-consistent poisoned samples stay invisible; ISN extracts a reusable antibody direction from the exposed ones and uses it to recover the hidden ones that share the same antigen.

Critical strengths
  1. Applying adaptive immunity to continual backdoor defense is conceptually interesting.
  2. The paper is clearly organized, and the progression from biological motivation to the ISN pipeline is easy to follow.
Critical weaknesses
  1. The paper does not discuss the case where the training stream contains only label-conflicting or only label-consistent poisoned samples, and such a discussion should be added.
  2. All experiments are conducted solely on a single frozen CLIP ViT-B/16 backbone, so it remains unclear whether the class-conditioned non-self residual comes from the backdoors themselves or from the stable semantic space of CLIP.
  3. The paper reports that backdoored samples consistently retain a recurring non-self residual after the clean response is subtracted, but does not explain why backdoors produce it or when this assumption fails.
Questions for authors

See Weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer f2QA #1

Dear Reviewer f2QA,

Thank you for recognizing the conceptual motivation and organization of the paper. The mechanism is to learn a recurring deviation from class-conditioned clean responses and use it to change poisoned supervision. We address the two extreme poison compositions, isolate the roles of the pretrained representation and reference calibration, and explain why residual sharing arises. New experiments use CIFAR-10 Task-IL and seed 42. CA is clean accuracy, and ASR is clean-conditioned unless explicitly called unconditional.

AuthorsOpenReview ↗

Response to Reviewer f2QA #2

W1. Entirely label-conflicting and entirely label-consistent streams

The discovery rule and the poison-label composition describe different aspects of the problem. Label composition concerns whether the observed label has been changed. Discovery concerns whether the reference provides enough reliable evidence to construct a detector. The submitted mechanism connects exposed and hidden examples within a stream. Appendix E.2, p. 24, states its evaluation question explicitly.

We examine whether antigen evidence exposed by semantically conflicting, cross-source poisoned samples transfers to label-consistent same-source backdoors.

When every poison is label-conflicting, correction supplies a direct response to the changed supervision. The exposed examples can also support antibody fitting and vaccination. Their presence does not require a corresponding population of label-consistent poisons. We have added both first-task and complete-stream controls in which all poisoned images come from a non-target class and are assigned the target label.

Poison compositionEvaluationVanilla CA (%)ISN CA (%)Vanilla ASR (%)ISN ASR (%)
Entirely label-conflictingFirst task99.8099.9592.480.00
Entirely label-conflictingFinal five-task test99.6999.6395.480.80
Mixed labelsFinal five-task test99.7399.6994.540.06

The first-task test starts with empty memory. ISN obtains 805 conflict seeds, including 800 poisons, and admits one local antibody. The five-task label-conflicting comparison changes conditional ASR from 4,754/4,979 for Vanilla to 40/4,969 for ISN. The mixed-label comparison changes it from 4,709/4,981 to 3/4,975. These within-condition controls establish protection both when all poisons have altered labels and when exposed and hidden poisons coexist.

When every poison retains the correct label, label agreement alone supplies no rejection signal. A compatible antibody can still flag a sample through abnormal affinity. Its neutralized training view keeps the semantic label, while vaccination places the learned factor on a non-target donor with that donor's label. This breaks the shortcut association through label-preserving supervision. Sec. 3.3, p. 6, makes the decision rule explicit.

Semantic consistency provides the retained label, and antibody evidence determines intervention.

For a first encounter, we have added a discovery supplement that proposes recurring local patterns from images and observed labels, verifies recurrence on another fold, and screens the candidates against clean references. It does not consume semantic-conflict seeds. We isolate this property using an effective fixed-patch experiment in which the target class naturally produces zero conflict seeds and neither class yields an accepted hypothesis. The poisons in this diagnostic are label-flipped. This allows us to control availability of discovery evidence separately from the labels' ground-truth correctness.

First-task condition with no usable class hypothesisCA (%)Unconditional ASR (%)Conditional ASR (%)
Clean ISN99.900.100.00
Poisoned ISN99.8092.0091.98
Poisoned ISN with class-matched random removal99.8091.9091.88
Poisoned ISN with revised screening99.950.100.00

Both removal arms delete 1,026 images, retain 8,654, and take 2,710 main optimizer updates. The screen removes all 968 poisons, while random removal removes 164. The revised route uses 120 distinct trusted examples per class in total. Its inputs and selection rule do not require labels to have been flipped, and the empirical result is reported under the actual label-flipped diagnostic. We will separate poison composition, semantic evidence, and the action taken on a flagged sample in Sec. 3.3.

AuthorsOpenReview ↗

Response to Reviewer f2QA #3

W2. The contribution of CLIP and the effect of reference errors

The defense gain persists when the pretrained backbone and task-learning architecture are held fixed. In the five-task label-flipping experiment, poisoned Vanilla reaches 95.48% conditional ASR and ISN reaches 0.80%, at 99.69% and 99.63% CA. Clean-trained Vanilla has 0.02% ASR under the same trigger. Thus, the frozen representation supports excellent clean classification while the task learner can still acquire the backdoor. The ISN mechanisms change that outcome under the shared architecture.

The clean reference is deliberately separated from the task expert. Sec. 3.2, p. 4, specifies:

The semantic reference always uses the unadapted f0f_0.

The paper defines prototypes from trusted images in Eq. (2), class-conditioned receptor centers and scales in Eq. (17), and residual aggregation in Eqs. (5)–(6). The spatial receptor observes RGB pixels, the frequency receptor observes DCT coefficients, and the model-conditioned receptor observes the temporary expert's activation-gradient contributions. CLIP provides the semantic coordinates used to condition those observations. The computation of repeated pixel or frequency structure is explicit in the receptor rather than supplied by a CLIP backdoor classifier.

The new anchor intervention isolates calibration from classification accuracy. We hold the poisoned stream, trigger, frozen encoder, and held-out cohorts fixed, replacing only the 20 anchors per class. The prototype and conflict-seed diagnostics take zero main-model optimizer steps.

Fixed-stream diagnosticIndependent proxy anchorsNative task anchors
Clean held-out semantic accuracy40/40 (100.00%)40/40 (100.00%)
Poisoned-source semantic accuracy965/968 (99.69%)966/968 (99.79%)
Poisons passing observed-label rejection3/968 (0.31%)773/968 (79.86%)
Poisons admitted as conflict seeds0/968 (0.00%)760/968 (78.51%)
Clean training conflicts5/8,712 (0.06%)5/8,712 (0.06%)

The semantic predictions differ on only one poisoned image, but the observed-label rejection criterion changes substantially. The native references restore 760 poison seeds while the number of clean conflicts stays at five. This identifies the reference distribution used by the empirical rejection test as the bottleneck in this case. High semantic classification accuracy and reliable seed coverage must therefore be evaluated separately.

We also perturb only the semantic branch with Gaussian smoothing, leaving the task-learning images unchanged. With the same 6,380-image training stream and 638 poisons, clean held-out semantic accuracy changes from 99.60% to 95.50%, poison-seed recall from 261/638 (40.91%) to 55/638 (8.62%), and seed precision from 261/263 (99.24%) to 55/81 (67.90%). This controlled intervention shows how reference degradation propagates into discovery and motivates both a stable frozen reference and the complementary acquisition route.

The downstream cost is measured separately. In the five-task label-flipping run, actual clean-label miscorrection is 68/44,460 (0.15%). In a fully clean stream it is 84/48,400 (0.17%), with final CA of 99.67% for ISN and 99.79% for Vanilla. These are actual label changes, rather than counting every flag as a wrong correction.

The experiments isolate the representation, empirical reference, and intervention within the CLIP-LoRA setting. We will describe the encoder as the fixed coordinate system used in this evaluation and keep the claim about residual extraction tied to the measured receptor evidence.

AuthorsOpenReview ↗

Response to Reviewer f2QA #4

W3. Why the residual recurs and when it becomes uninformative

The design separates a recurring attack-related component from the normal variation of the source content. A useful local model for a poisoned observation is

\approx b_{r,q,c}+\delta_{r,q,a}+\xi_x,$$ where $b_{r,q,c}$ is the expected clean response for content class $c$, $\delta_{r,q,a}$ is a component recurring under attack realization $a$, and $\xi_x$ captures remaining input variation. Subtracting and scaling the clean response as in Eq. (5) gives $$\rho_{r,q}(T_a(x);c) \approx P_r\!\left( \frac{\delta_{r,q,a}+\xi_x}{\sigma_{r,q,c}+\varepsilon} \right).$$ This is an explanation of the working assumption. A recurring local patch can produce consistent pixel deviations, a repeated warp can alter the distribution of target-conditioned contributions, and a frequency perturbation can recur at transform coordinates. Class conditioning removes the expected content response before aggregation across seeds estimates what repeats. The controls in Appendix E.2 directly test the value of that subtraction under a fixed encoder. | Attack and receptor | Random-direction AUROC | Raw-template AUROC | Self-residual AUROC | Frozen clean FPR (%) | Full-stream same-source recall (%) | | --- | ---: | ---: | ---: | ---: | ---: | | BadNets, spatial | 0.489 | 0.558 | **0.976** | 7.2 | 93.2 | | WaNet, model-conditioned | 0.543 | 0.843 | **0.884** | 2.9 | 44.8 | | FTrojan, frequency | 0.245 | 0.478 | **1.000** | 6.7 | 91.6 | The raw-template and random-direction controls use the corresponding observation spaces. For WaNet, model confidence itself has AUROC 0.420, below the residual receptor's 0.884. Appendix E.2, p. 25, gives the interpretation of its raw-contribution control: > *The gain over the raw-contribution control isolates the benefit of removing normal class-conditioned model response rather than relying on the expert’s activation pattern alone.* The AUROC columns measure ranking among same-source candidates. The final two columns measure operation at the frozen threshold and include the effect of candidate routing on full-stream recall. Appendix E.2, p. 26, summarizes the common observation: > *In each case, an antibody fitted from semantic-conflict evidence produces an affinity response on hidden, label-consistent same-source candidates.* **The latest source-stage validation also tests a refined spatial representation under the same admission cutoffs.** The representation selects a connected component from the learned mask, corrects background using its one-pixel exterior, and fits spatial statistics and energy-based scores. Candidate construction uses the learned support and source data. Source fitting and validation folds are separate. | Source task | Original direction recall | Revised representation recall | Revised AUC | Revised effect size | | --- | ---: | ---: | ---: | ---: | | Task 0 | 13/283 (4.59%) | **176/283 (62.19%)** | 0.9524 | 2.2926 | | Task 1 | 101/177 (57.06%) | **146/177 (82.49%)** | 0.9499 | 1.7788 | These metrics use held-out source seeds. Both revised candidates satisfy the original AUC, effect-size, recall, and support thresholds. Their supports of 863 and 1,388 consolidate into one mature spatial cell with two source-task supports and total support of 2,251. This provides a concrete source-stage example of extracting a reusable spatial factor through a revised observation rule while retaining the admission criteria. The recurrence assumption is evaluated through those acceptance tests. An informative component separates from legitimate class variation and remains identifiable under the chosen content reference. Sec. 3.3, p. 6, states: > *An antibody is locally validated only when it separates held-out seeds from content-matched trusted-clean examples under fixed ranking, effect-size, and low-false-positive criteria.* The default tests require AUC of at least 0.65, normalized separation of at least 0.5, and validation recall of at least 0.1. Response views must additionally preserve the semantic class and satisfy the directional affinity tests in Eqs. (8) and (27). The abstraction therefore has observable acceptance criteria. We will replace the introduction's unconditional wording with the measured phenomenon of recurring class-conditioned residuals and state these acceptance conditions alongside it. Thank you again for the questions. The composition controls, reference interventions, and receptor-matched comparisons make the source of the defense gain and the residual assumption explicit.
Reviewer_DSaUOpenReview ↗
Rating
3
Confidence
4(+1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes ISN, a backdoor defense for task-incremental learning. A frozen semantic encoder identifies label-conflicting examples, whose class-conditioned residuals are aggregated into validated “antibodies.” These guide label correction, neutralization, vaccination, and cross-task memory. Experiments on CIFAR-10, CIFAR-100, and TinyImageNet report strong attack suppression while preserving clean accuracy under BadNets, WaNet, and LTB.

Critical strengths
  1. A useful mechanism beyond semantic disagreement. The residual-based detectors recognize label-consistent poisons using evidence from label-conflicting examples. The raw-template and random-direction controls in Appendix E.2 help substantiate this mechanism.
  2. Strong controlled evaluation. The matched CLIP–LoRA comparisons show a favorable accuracy–security trade-off across three datasets. Separating antibody fitting, validation, and final calibration, together with the component ablations, strengthens the empirical evidence.
Critical weaknesses
  1. W1 — Limited cold-start protection. Equation (4) requires sufficiently many semantic conflicts to discover a new antibody. As Appendix B acknowledges, an unseen attack producing too few such examples may cause abstention unless compatible memory exists. Appendix E.2 tests transfer within an attack, not end-to-end defense against a novel purely label-consistent attack. This materially limits the claimed protection against hidden infections.
  2. W2 — Cross-family memory is not established. The abstract claims reusable memory across attack families, whereas the main streams repeat BadNets or WaNet across tasks. Table 5 supports memory reuse for BadNets, but does not establish cross-family transfer. Moreover, the reported 77.1% late-task reduction corresponds to only 0.0219% to 0.0050% ASR (Figure 3). Claims should distinguish same-attack reuse, changed-trigger transfer, and cross-family generalization.
  3. W3 — Trusted-data requirements constrain the operating range. The method reserves 20 anchors plus 20 calibration examples per class—8% of training data on CIFAR-100 and TinyImageNet. With 20 anchors, semantic scores cannot fall below 1/211/21. Appendix E.6 consequently shows approximately 94%/62% ASR for BadNets/WaNet when the shared level is at most 0.025. This disclosed boundary is not a mathematical error, but leaves smaller clean budgets and stricter false-positive operating points unsupported.
Questions for authors
  1. Does any reported end-to-end evaluation contain a previously unseen, purely label-consistent attack with neither semantic-conflict examples nor compatible memory? If not, should the corresponding claim be restricted to transfer from attacks exposing semantic conflicts?
  2. Which reported experiment demonstrates memory transfer across changed triggers or attack families? From existing logs, can you distinguish detections contributed only by recalled memory from those already covered by current-task antibodies?
  3. Can semantic-discovery and final-calibration levels be decoupled? What minimum trusted-data budget would permit a nominal 1% screening level without forced abstention?
4 comments
AuthorsOpenReview ↗

Response to Reviewer DSaU #1

Dear Reviewer DSaU,

Thank you for recognizing the residual-based mechanism beyond semantic disagreement and the controlled evaluation. ISN transfers evidence from exposed examples to label-consistent poisons and retains validated detectors for recurrent threats. We address antigen acquisition, the contribution of immune memory, and the trusted-data requirements below, using both the submitted formulation and the new controls. Supplementary experiments use CIFAR-10 Task-IL and training seed 42. CA is clean accuracy. Conditional ASR uses triggered non-target examples whose clean versions are correctly classified, while unconditional ASR uses all triggered non-target examples.

AuthorsOpenReview ↗

Response to Reviewer DSaU #2

W1 / Q1. Acquiring the first antigen evidence

We directly test the acquisition failure you identify and show that a complementary discovery route restores protection under an effective attack. The distinction is between having no historical memory and having no usable current discovery hypothesis. Our experiments now examine both.

The submitted mechanism is stated in Sec. 3.3, p. 4.

Immune perception operationalizes this evidence transfer in three steps: it exposes cross-class seeds, fits their recurring non-self response through complementary receptors, and uses validated antibodies to screen same-class data.

This makes the role of label consistency precise. A poisoned example can agree with its observed label and still express the residual learned from another example. The exposed subset supplies the recognition rule, and the label-consistent subset is screened using affinity. Sec. 3.3, p. 6, states the separation between the two decisions.

Semantic consistency provides the retained label, and antibody evidence determines intervention.

First-task protection with empty memory. In the all-label-flipping BadNets experiment, each task contains 988 poisoned examples among 9,880 training images. The first task starts with an empty memory bank. Poisoning raises Vanilla's conditional ASR from 0/1,000 to 922/997, while ISN gives 0/1,000 at 99.95% CA.

First-task training conditionCA (%)Unconditional ASR (%)Conditional ASR (%)
Clean Vanilla LoRA99.950.000.00
Poisoned Vanilla LoRA99.8092.5092.48
Poisoned ISN99.950.000.00

ISN obtains 805 semantic-conflict seeds, of which 800 are poisoned, admits one local antibody, and recalls zero memory antibodies. These counts establish that prior exposure through memory is unnecessary when the current task supplies reliable evidence.

Natural absence of a usable discovery hypothesis. We then use a fixed reference split under which the target class has zero semantic-conflict seeds. The other class produces five clean conflicts, below the minimum support of six, so neither class supplies an accepted hypothesis. The first task contains 968 label-flipping poisons among 9,680 images, and the original ISN acquires the backdoor. The measured attack therefore exercises the acquisition failure directly.

We add a complementary acquisition-and-screening route. It proposes repeated local patterns from training images and observed labels, checks recurrence on a separate training fold, and uses trusted clean references to screen out common clean patterns. Candidate construction uses no semantic-conflict seed input. The retained stream is then passed to the ISN learner. Poison identities enter only the retrospective audit.

First-task conditionCA (%)Unconditional ASR (%)Conditional ASR (%)
Clean ISN99.900.100.00
Poisoned ISN99.8092.0091.98
Poisoned ISN with class-matched random removal99.8091.9091.88
Poisoned ISN with revised screening99.950.100.00

Screening and random removal each delete 1,026 images under the same observed-class quotas, retain 8,654 images, and perform 2,710 main optimizer updates. Screening removes all 968 poisons and 58 clean examples. Random removal removes 164 poisons and leaves 804. The 91.80-percentage-point unconditional ASR difference therefore comes from which examples are selected under the matched deletion and update budgets.

The experiment uses 120 distinct trusted examples per class in total, including 60 additional examples beyond the three-split supplementary implementation. Its poisons are label-flipped and its target-class seed set is naturally empty. These two properties are reported separately because poison-label correctness and passage through the empirical discovery gate are different quantities.

For a wholly label-consistent stream, discovery must draw on recurring evidence in the images or a compatible previously validated detector. The complementary route searches for the former without requiring a disagreement with the observed label. We will describe the new result as acquisition under naturally absent usable target seeds and retain exposed-to-hidden transfer as the claim tested in Appendix E.2.

AuthorsOpenReview ↗

Response to Reviewer DSaU #3

W2 / Q2. What transfers through memory and what it contributes

Memory retains validated residual patterns and recurrence evidence. In class-disjoint streams, current-task labels, clean centers, and thresholds are rebuilt before reuse. Appendix D.7, p. 22, states:

Current anchors rebuild class centers and scales.

The stored detector operates against the new semantic reference. Appendix D.7 also permits spatial and frequency recall without new local seeds.

If such evidence is unavailable or the test fails, the cell is excluded for that label; spatial and frequency recall can still operate without a new local seed set.

"Such evidence" refers to model-conditioned cells' current-seed validation. Spatial and frequency cells reuse stored coordinates with current clean calibration.

The new component comparison freezes source memory and changes the query pattern. The first two tasks establish a white-patch history. The query uses a checkerboard at the same position and size. The revised source-fitted spatial representation passes the original validation cutoffs on both source tasks and supplies a mature cell with two distinct task supports. At the query stage, semantic-conflict seed use, local receptor fitting, heating, and vaccination are disabled. The two ISN arms share the source experts and frozen bank and differ in whether historical memory can act.

Query condition, training holdoutCA (%)Unconditional ASR (%)Conditional ASR (%)
Clean Vanilla99.900.000.00
Poisoned Vanilla99.70100.00100.00
Poisoned ISN, memory disabled99.70100.00100.00
Poisoned ISN, revised spatial memory99.800.200.00

The completed query arms each use 8,880 images, ten epochs, and 2,780 main optimizer updates. The poisoned arms contain 888 poisons. Enabling memory reduces unconditional ASR from 500/500 to 1/500, a 99.80-point reduction, and raises CA by 0.10 points. In the poisoned training stream, detection recall is 888/888, clean false flags are 95/7,992 (1.19%), and precision is 888/983 (90.34%).

All 50 parameter tensors of the first two experts retain their source checkpoint values. Stored memory contents and the complete bank fingerprint also remain unchanged through the query. Current clean references calibrate the frozen detector, isolating historical evidence under the pattern change.

Training-holdout component results from four completed arms. The unfinished clean-ISN query control leaves the five-arm acceptance protocol incomplete. The query test set was not evaluated.

The earlier logs distinguish extra memory hits from current-antibody coverage. In the supplementary mixed-label stream, at the common joint threshold, recalled memory flags 8/494 label-consistent poisons missed by current antibodies, together with 100/8,892 clean examples. Counts are per unique image.

We additionally compare separately calibrated local-only and local-plus-memory views from the same trained model in the label-flipping stream. Task indices are zero-based.

TaskLocal poison detectionsLocal plus memory poison detectionsLocal clean flagsLocal plus memory clean flags
282/98816/988319/8,892392/8,892
3976/988976/988243/8,89249/8,892
4957/988958/98890/8,89290/8,892

At task 3, recall preserves 976/988 detections while avoiding 194 clean flags, a reduction from 2.73% to 0.55%. The task-2 row shows why calibration belongs in the definition of recall. Adding another detector changes the maximum-score distribution and its threshold, so the operational result depends on both coverage and clean-reference calibration. These readouts measure the decisions of a fixed trained model. Final ASR also depends on the subsequent correction, neutralization, and vaccination routes.

The method explicitly prevents recall from manufacturing its own supporting evidence. Appendix D.7, p. 21, states:

Thus an antibody cannot create its own admission or maturation evidence through repeated activation.

We will distinguish cross-source transfer, compatible recurrence across tasks, and use of source-fitted memory after a pattern change. The new component study tests the third operation on a training holdout, and the earlier memory-hit counts report incremental screening decisions. The abstract will identify these operations explicitly. For Fig. 3(c), we will show the absolute change from 0.0219% to 0.0050%, or 0.0169 percentage points, alongside the relative percentage. These comparisons quantify memory's contribution.

AuthorsOpenReview ↗

Response to Reviewer DSaU #4

W3 / Q3. Trusted-data requirements and independently chosen screening levels

Discovery and final screening can use separate levels. Equation (4) governs exposure of seed evidence, while Eq. (23) calibrates the final maximum-affinity score after the acting detector set is fixed. The shared-level sweep in Appendix E.6 changes both stages simultaneously.

For nanchorn_{\mathrm{anchor}} anchors, Eq. (16) has the lower bound

pq(x)≥1nanchor+1.p_q(x)\geq \frac{1}{n_{\mathrm{anchor}}+1}.

With 20 anchors, this is 1/21≈0.04761/21\approx0.0476. A rejection level of 0.01 or 0.025 therefore cannot expose a seed through that rule, regardless of how atypical the sample appears. Keeping the discovery level at 0.05 preserves the original resolution while allowing a stricter final screening level.

For the final calibration operator in Appendix D.5, p. 19, a finite threshold requires

\leq n_{\mathrm{cal}}, \qquad \alpha_{\mathrm{screen}}\geq \frac{1}{n_{\mathrm{cal}}+1}.$$ At a nominal 1% screening level, the first finite rank occurs with 99 calibration examples per class. Retaining 20 anchors gives 119 trusted examples per class in the paper's two-subset protocol. Retaining the supplementary implementation's separate 20-example validation set gives 139. If discovery itself also uses 1%, at least 99 anchors are required for that rejection level. Joint calibration is also important when local, recalled, and newly acquired detectors coexist. Appendix D.5, p. 19, explains: > *Calibrating the maximum directly accounts for the current and recalled antibodies in the same decision and avoids inflating the clean false-positive rate through an independent OR rule.* We will expose discovery, validation, and final screening as separate operating choices. We will also report the trusted-data accounting with each experiment. The paper's two-split setting uses 40 examples per class, which is 8% for classes with 500 training images. The supplementary three-split setting uses 60 examples per class, and the acquisition result above uses 120 distinct examples per class after accounting for overlap with its screening calibration pool. The 99-example calculation specifies the final threshold's reference resolution, while the empirical tables specify the performance and budget actually evaluated. Thank you again for the careful questions. The new acquisition control, explicit memory accounting, and separate calibration levels make the mechanism and its operating conditions directly inspectable.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.