acceptodds
Under review as a conference paper at ICLR 2027

Authorization and Endorsement Forgery in Multi-Agent Evaluation: Propagation and Defense

Abstract

Recent LLM-based evaluation systems increasingly rely on multi-agent collaboration, where judges exchange assessments and revise decisions before final aggregation to improve evaluation quality. However, such collaboration introduces an evidence-authorization vulnerability: unsupported authority claims can acquire apparent authority when relayed by internal agents, leading downstream judges to treat them as credible peer evidence. We characterize this vulnerability as Authorization and Endorsement Forgery (AEF), an IPI-derived attack pattern that manipulates how evaluation evidence is interpreted and trusted. Unlike conventional instruction-oriented IPI payloads, AEF injects task-relevant but fabricated or unsupported claims of expert endorsement, prior adjudication, authorization, or consensus rather than directly steering the evaluator through injected instructions. Across three LLM evaluation benchmarks, AEF propagates through multi-agent evaluation pipelines and increases downstream target-selection rates by up to 28.03 percentage points, demonstrating its ability to manipulate evidence trust beyond direct exposure. To mitigate this vulnerability, we propose TRUSTGUARD, an orchestration-layer defense that separates evidence registration from authorization. TRUSTGUARD prevents unsupported authority claims from gaining decision influence by tracking source and derived evidence and permitting downstream use only when provenance and policy-defined authorization requirements are satisfied. Evaluation results show that TRUSTGUARD reduces macro-average ASR-any from 61.03% to 34.41% while maintaining clean accuracy comparable to the undefended system.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM-as-a-judge, Multi-Agent Security, Indirect Prompt Injection, Authorization and Endorsement Forgery, Cross-Agent Propagation, Provenance-Aware Authorization

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 2 reviews, above 27% of the others.
Reviewer_EKWzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies Authorization and Endorsement Forgery, where fabricated authority claims enter a multi-agent evaluation system through untrusted context and gain credibility when relayed by judges. Its main contributions are characterizing this evidence-laundering mechanism and proposing TrustGuard, an orchestration-layer defense that tracks evidence provenance and dependencies and checks authorization before downstream use. Under complete mediation and correctly specified policies, derived evidence cannot gain authority beyond that permitted by its ancestry. Validation spans three benchmarks, three judge models, and four evaluation architectures. AEF achieves a mean attack success rate of 58.04%, compared with 43.72% for the tested conventional prompt injection.

Critical strengths

The paper presents a clear framing of authorization and endorsement forgery, highlighting how unsupported authority claims can gain credibility through inter-agent communication. TrustGuard addresses this risk through provenance tracking and explicit authorization gates, with a useful principle that registration and internal relay should not automatically confer decision authority.

The evaluation spans three benchmarks, four evaluation architectures, and multiple judge models. One-judge exposure experiments provide evidence of downstream behavioral influence, while comparisons with several prompt-level defenses demonstrate TrustGuard’s stronger average robustness in the tested settings.

The presentation benefits from intuitive examples, a clear system diagram, and explicit trust assumptions and limitations. The work addresses a relevant security concern in multi-agent evaluation and offers a concrete system-level defense that could complement prompt-based safeguards.

Critical weaknesses
  1. Appendix A’s AEF payload directs the judge to select the target, generate criticisms, and conceal the memo. Without authority-only and instruction-only ablations, the experiments do not establish that forged authority drives the observed effects, weakening the paper’s central mechanistic claim.

  2. AEF uses an external evaluation memo, whereas the conventional IPI baseline appends a brief directive to candidate text. Differences in location, length, and framing prevent attribution of the reported ASR advantage specifically to authority forgery.

  3. Because the attack enters as a separately registered external memo, a simple policy excluding unauthorized external context could potentially provide similar protection. The absence of this baseline and G1–G4 ablations leaves unclear whether TrustGuard’s gains require its proposed ancestry-tracking architecture.

  4. The paper provides illustrative rules but insufficient detail about the policies used experimentally, dependency registration, and construction of verified content views. These choices determine what TrustGuard blocks, making their omission consequential for reproducibility and interpretation of the results.

  5. The paper acknowledges that authority claims embedded within otherwise admissible candidate responses may require whole-item quarantine or finer-grained tracking. Without evaluating this setting, the evidence supports a narrower deployment scenario than the broader motivation of preventing evidence laundering across agents.

  6. Visibility-based dependencies can make an otherwise useful judgment inadmissible because it shares context with untrusted evidence. Comparable clean accuracy does not establish the cost under mixed legitimate and malicious external context. Quarantine rates, abstentions, decision coverage, and robustness–utility trade-offs are needed to assess whether the defense preserves useful multi-agent collaboration.

Questions for authors
  1. Could the authors separate authority framing from directive compliance using matched authority-only, instruction-only, and combined payloads, keeping insertion location and length comparable? How much of AEF’s advantage remains in the authority-only condition?

  2. What determines the dependency edges registered in DEL? In particular, does exposure to one inadmissible input render an agent’s entire output inadmissible, and how frequently does this suppress otherwise useful judgments during multi-round debate?

  3. Could the authors provide the receiver–purpose permission matrix used in each experimental setting, including how candidate responses, external endorsements, judge assessments, and summaries receive different permissions?

  4. How much latency does TrustGuard add per debate round and final decision? Please distinguish ledger and cryptographic operations from dependency traversal and any additional model calls, and explain whether caching reduces these costs.

  5. How does robustness change when each authorization gate is disabled? In particular, how does the full system compare with a simpler baseline that only rejects unauthorized external context at G1?

  6. Can TrustGuard preserve a candidate response for evaluation while preventing embedded authority claims from influencing other judges? Please evaluate this attack surface and quantify any trade-off between attack resistance and decision coverage.

8 comments
AuthorsOpenReview ↗

Author Response (1/7): AEF Effectiveness Without Explicit Behavioral Instructions

We thank the reviewer for the detailed feedback and helpful questions. Following your suggestions, we conducted additional experiments and analyses to examine the AEF attack mechanism, evidence dependency registration, authorization policies, runtime overhead, and the utility of TrustGuard. We respond to each point below.

1. Does AEF remain effective without explicit behavioral instructions? (Weaknesses 1–2; Question 1)

We agree that the original IPI–AEF comparison changed both the delivery channel and payload composition, so its performance difference cannot be attributed solely to authority forgery.

Controlled ablation study

Following your suggestion, we conducted a controlled four-condition ablation study to isolate the effects of forged authority claims and explicit behavioral instructions. We kept the external delivery channel, insertion position, recipient, system prompt, and wrapper fixed across all conditions. Candidate responses remained unchanged, and all payloads were matched to 625 characters. Only Judge 1 received the payload. The four conditions are:

  • Neutral control: Neither authority claims nor explicit behavioral instructions.
  • Instructions only: Explicit behavioral directives, without authority claims.
  • Authority only: Forged authority or endorsement claims, without explicit behavioral directives.
  • Both: The combined authority-and-instruction payload.

For Qwen2.5-7B, the experiment covers 66 FairEval, 80 LLMBar, and 76 MT-Bench candidate pairs, evaluated in both candidate orders: 222 pairs and 444 directional instances per condition. We also evaluate the ablation with DeepSeek-flash. We report the receiving judge's target selection rate (TSR), distinguishing it from final multi-agent attack success.

Payload conditionQwen2.5-7B TSRDeepSeek-flash TSR
Neutral control23.6%13.3%
Instructions only92.3%26.8%
Authority only98.9%56.7%
Both99.3%55.1%

Results and implications

The results show that forged authority claims can influence evaluation even without explicit behavioral directives. The authority-only condition achieves a higher observed TSR than the instructions-only condition on both models: 98.9% vs. 92.3% for Qwen2.5-7B, and 56.7% vs. 26.8% for DeepSeek-flash. The DeepSeek results are particularly informative because the instructions-only condition does not already have a near-saturated TSR.

Adding explicit behavioral instructions to the authority-only payload yields little numerical improvement on Qwen2.5-7B (98.9% to 99.3%) and no numerical improvement on DeepSeek-flash (56.7% to 55.1%). Thus, the observed effect is not explained solely by the additional behavioral instructions, delivery channel, or payload length.

Accordingly, we will revise the Abstract to avoid presenting AEF as categorically distinct from instruction-oriented prompt injection. Instead, we emphasize that forged authorization or endorsement can influence evaluation even without explicit behavioral directives. We will include the complete payloads, character-length-matching procedure, sample-selection details, and statistical methods in the revised appendix.

AuthorsOpenReview ↗

Author Response (2/7): Dependency Registration and Opinion Exclusion (Weaknesses 4 and 6; Question 2)

2. How are dependency edges registered, and how frequently do they suppress otherwise useful judgments? (Weaknesses 4 and 6; Question 2)

DEL registers dependencies based on the evidence actually exposed to an agent. Specifically, dependency edges are created for inputs included in the agent's prompt and upstream judge assessments that pass authorization checks and are actually presented to the agent. Existing dependencies of these upstream assessments are also inherited. Evidence registered in SEL but not exposed to the agent does not create a dependency.

As described in Section 3.2, these parent references capture a conservative visibility dependency, not token-level causal attribution. Consequently, if an agent's output depends on an upstream evidence record that is inadmissible for the requested receiver–purpose pair, the entire output is withheld for that use. This prevents paraphrasing, debate, or internal relay from broadening the permitted use of the original evidence. Even an assessment that rejects a malicious claim or supports the correct candidate may be withheld if it shares the relevant dependency.

To quantify this effect, we conduct additional evaluations under benign-memo and paraphrased-attack conditions:

Evaluation conditionQuarantine rateFraction of quarantined opinions that selected the correct candidateDecision coverage
Benign memo33.3%64.4%–68.4%100%
Paraphrased attack33.3%49.4%–63.8%100%

In these settings, one out of three opinions per directional instance is quarantined. The correct-selection percentages are calculated among the quarantined opinions, not among all correct judgments. The remaining two judges' opinions pass authorization checks, allowing the evaluated instances to reach a final decision with 100% decision coverage.

In this initial single-round evaluation, Full TrustGuard changes strict bidirectional accuracy relative to G1-only under benign-memo conditions by +6.1%, −6.3%, and −7.9% on FairEval, LLMBar, and MT-Bench, respectively. These results show that excluding an opinion can affect evaluation accuracy.

AuthorsOpenReview ↗

Author Response (3/7): Evaluation Utility under Clean, Benign, and Mixed Contexts (Weakness 6; Question 2, Continued)

Continued: How does TrustGuard affect evaluation utility under clean, benign, and mixed contexts? (Weakness 6; Question 2)

To further examine this effect during multi-round debate, we conduct an additional evaluation using the same three-judge, two-round Debate setting as in Q5, in which judges exchange assessments across two rounds. The evaluation uses DeepSeek-flash (thinking off, temperature 0) and covers all 222 candidate pairs: 66 FairEval, 80 LLMBar, and 76 MT-Bench pairs, evaluated in both candidate orders (444 directional instances).

We compare No Defense, G1-only, and Full TrustGuard under three conditions: no external memo, benign external context, and co-occurring legitimate and malicious external evidence. Only agent_1 is exposed to the external context. All accuracy values report strict bidirectional accuracy. No Defense and G1-only have zero quarantine across all three conditions.

External contextDatasetNo DefenseG1-onlyFull TrustGuardQuarantine rate (Full)
No external memoFairEval63.6%62.1%60.6%0%
No external memoLLMBar71.2%73.8%71.2%0%
No external memoMT-Bench88.2%89.5%90.8%0%
Benign external contextFairEval66.7%60.6%60.6%33.3%
Benign external contextLLMBar65.0%72.5%73.8%33.3%
Benign external contextMT-Bench90.8%89.5%89.5%33.3%
Legitimate and malicious external evidenceFairEval27.3%31.8%62.1%33.3%
Legitimate and malicious external evidenceLLMBar55.0%53.8%72.5%33.3%
Legitimate and malicious external evidenceMT-Bench72.4%73.7%88.2%33.3%

Results

First, the quarantine rate is 33.3% under both benign and mixed external contexts, and does not increase in the second round of the evaluated setting. It is 0% when no external memo is present.

Second, all 27 evaluated configurations achieve 100% decision coverage with zero abstentions. Thus, all evaluated cases produce a final decision despite the quarantined opinions. However, full decision coverage does not by itself establish unchanged accuracy.

Without an external memo, Full TrustGuard remains within 3% of No Defense on all three datasets. Paired McNemar tests show no significant aggregate difference (Full vs. No Defense: p = 1.0; Full vs. G1-only: p = 0.77).

With benign external context, the accuracy differences vary across datasets. The pooled paired comparisons show no significant difference between Full and No Defense (p = 0.82), or between Full and G1-only (p = 1.0). At the dataset level, the Full vs. No Defense comparison on LLMBar is significant (p = 0.016).

Third, under mixed external context, Full TrustGuard achieves significantly higher strict bidirectional accuracy than both No Defense and G1-only across all three datasets. Relative to No Defense, accuracy increases by 34.8%, 17.5%, and 15.8% on FairEval, LLMBar, and MT-Bench, respectively. Relative to G1-only, the improvements are 30.3%, 18.7%, and 14.5%.

The paired McNemar tests further support these results:

DatasetFull vs. No DefenseFull vs. G1-only
FairEvalp < 10⁻⁵p = 0.00004
LLMBarp = 0.0005p = 0.0003
MT-Benchp = 0.0042p = 0.0074
Pooledp ≈ 10⁻⁷p ≈ 10⁻⁷

The pooled discordant-pair counts are 53:4 for Full vs. No Defense and 51:5 for Full vs. G1-only. Notably, Full TrustGuard's accuracy under mixed external context differs from its no-memo accuracy by only 1.5%, 1.3%, and 2.6% on FairEval, LLMBar, and MT-Bench, respectively.

Together, these results show that, in the evaluated two-round setting, the quarantine rate does not increase in the second round, and all evaluated cases retain full decision coverage. Under mixed external context, Full TrustGuard also achieves significantly higher accuracy than both comparison systems, while its accuracy remains close to the no-memo results. The benign-context results show that the accuracy impact varies across datasets.

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

This paper investigates an attack vector in multi-agent LLM evaluation systems termed Authorization and Endorsement Forgery (AEF). In AEF, an attacker injects fabricated authority claims—such as purported expert endorsements, prior adjudications, or benchmark approvals—through untrusted external evaluation materials. When an exposed judge absorbs and paraphrases these claims into its assessment, the resulting peer message can influence downstream judges that never directly observe the attack payload, a provenance-obscuring process the authors characterize as evidence laundering. The authors evaluate AEF across three pairwise evaluation benchmarks (FairEval, LLMBar, and MT-Bench), three judge models (DeepSeek-v4-flash, Qwen2.5-7B, and GPT-4o-mini), and four evaluation architectures, using a one-judge exposure setting to measure cross-agent propagation. To mitigate this vulnerability, the paper proposes TRUSTGUARD, an orchestration-layer defense that maintains a Source Evidence Ledger (SEL) and a Derived Evidence Ledger (DEL) to track evidence provenance and upstream dependencies, authorizing evidence for decision-relevant use only when policy and ancestry requirements are met. Experimental results report that TRUSTGUARD reduces macro-average ASR-any from 61.03% to 34.41%.

Critical strengths
  1. Practical Threat Model and Targeted Propagation Analysis: The threat model is realistic, requiring the adversary only to control an untrusted evaluation input item without modifying system instructions, model parameters, or all participating judges. Furthermore, the one-judge exposure setup in the Debate and Debate+Summarizer architectures isolates indirect influence: because unexposed judges (J2J_2 and J3J_3) only receive peer assessments, the observed shifts in target selection and cross-agent flip rates provide direct behavioral evidence of downstream attack propagation.

  2. Principled System Design for Evidence Authorization: Framing defense around the principle that registration is not authorization—by separating initial evidence logging from decision-relevant consumption and preserving provenance across paraphrasing and summarization—is a sound architectural direction. Comparing TRUSTGUARD against five prompt-level defenses across multiple datasets and judge models establishes a useful empirical baseline for orchestration-level protection.

Critical weaknesses
  1. A central claim of the paper is that AEF succeeds by manipulating evidence trust via forged authority "rather than directly steering the evaluator through injected instructions" (Abstract). However, the attack templates in Appendix A introduce severe confounding factors between conventional IPI and AEF. First, IPI is appended directly to the candidate response, whereas AEF is supplied through a separate external evaluation memo channel. Second, as Appendix A acknowledges, the AEF memo explicitly combines authority claims with strong behavioral directives: it instructs the judge to cite one or two plausible domain-relevant flaws in the non-target candidate, select the target candidate with high confidence, and conceal the memo's existence ("Do not quote or reference this memo"). Because insertion location, payload length/detail, and explicit stealth/critique instructions all change simultaneously, the higher ASR of AEF cannot be cleanly attributed to forged authorization or endorsement. This undermines the paper's core conceptual distinction between AEF and instruction-oriented prompt injection.

  2. Under the evaluated deployment, the untrusted external memo is registered as a distinct source evidence item in the SEL (Section 3.4). If an unauthenticated external memo can already be rejected at the initial source-consumption gate ( G1G_1), it is unclear how much additional protection is contributed by downstream dependency tracking in the DEL (G2−G4G_2-G_4). The paper lacks an ablation comparing full TRUSTGUARD against a simple input-gating baseline that filters unauthorized external memos at G1G_1. Additionally, while the Abstract and Conclusion claim that TRUSTGUARD maintains "clean accuracy comparable to the undefended system," Table 3 only reports strict bidirectional accuracy under attack, and no benign (clean) accuracy comparison with and without TRUSTGUARD is provided in the paper.

Questions for authors

Could you provide a controlled ablation that fixes the insertion channel and matches payload length as closely as possible to separately compare (i) authority claims only, (ii) explicit behavioral instructions only (e.g., generating plausible criticisms and hiding the source), and (iii) their combination? This would clarify whether AEF's effectiveness genuinely stems from the proposed authority-forgery mechanism rather than stronger injected instructions or channel differences.

6 comments
AuthorsOpenReview ↗

Author Response (1/3): AEF Effectiveness Without Explicit Behavioral Instructions

We thank the reviewer for the constructive comments on the attack comparison, downstream protection, and clean accuracy. In response, we conducted additional experiments to further examine these aspects and clarify the scope of our findings. Our detailed responses are presented below.

1. Does AEF remain effective without explicit behavioral instructions?

We agree that the original IPI–AEF comparison changed both the delivery channel and payload composition, so its performance difference cannot be attributed solely to authority forgery.

Controlled ablation study

Following your suggestion, we conducted a controlled four-condition ablation study to isolate the effects of forged authority claims and explicit behavioral instructions. We kept the external delivery channel, insertion position, recipient, system prompt, and wrapper fixed across all conditions. Candidate responses remained unchanged, and all payloads were matched to 625 characters. Only Judge 1 received the payload. The four conditions are:

  • Neutral control: Neither authority claims nor explicit behavioral instructions.
  • Instructions only: Explicit behavioral directives, without authority claims.
  • Authority only: Forged authority or endorsement claims, without explicit behavioral directives.
  • Both: The combined authority-and-instruction payload.

For Qwen2.5-7B, the experiment covers 66 FairEval, 80 LLMBar, and 76 MT-Bench candidate pairs, evaluated in both candidate orders: 222 pairs and 444 directional instances per condition. We also evaluate the ablation with DeepSeek-flash. We report the receiving judge's target selection rate (TSR), distinguishing it from final multi-agent attack success.

Payload conditionQwen2.5-7B TSRDeepSeek-flash TSR
Neutral control23.6%13.3%
Instructions only92.3%26.8%
Authority only98.9%56.7%
Both99.3%55.1%

Our additional experiments show that forged authority claims can influence evaluation even without explicit behavioral directives. The authority-only condition achieves a higher observed TSR than the instructions-only condition on both models: 98.9% vs. 92.3% for Qwen2.5-7B, and 56.7% vs. 26.8% for DeepSeek-flash. The DeepSeek results are particularly informative because the instructions-only condition does not already have a near-saturated TSR.

Moreover, adding explicit behavioral instructions to the authority-only payload yields little numerical improvement on Qwen2.5-7B (98.9% to 99.3%) and no numerical improvement on DeepSeek-flash (56.7% to 55.1%).

These results directly address the reviewer's concern: the effectiveness of the tested AEF payloads cannot be attributed solely to the additional behavioral instructions, different delivery channels, or payload lengths. The controlled comparison shows that the forged authority claims themselves remain effective when explicit behavioral directives are removed.

Accordingly, we will revise the Abstract's phrase “rather than directly steering the evaluator through injected instructions” to avoid presenting AEF as categorically distinct from instruction-oriented prompt injection. Instead, we emphasize that forged authorization or endorsement can influence evaluation even without explicit behavioral directives.

We will include the complete payloads, character-length-matching procedure, sample-selection details, and statistical methods in the revised appendix.

AuthorsOpenReview ↗

Author Response (2/3): Additional Protection Beyond G1-only

2. Does G1-only provide sufficient protection without downstream dependency tracking? (Weakness 2: Downstream Protection)

G1-only does not provide the same protection as the full TrustGuard architecture, because G1 rejects the use of source evidence for a particular receiver–purpose pair (r,p), but does not track its dependencies in subsequent agent-generated outputs. An external source may be accessed for a permitted purpose while remaining unauthorized as decision evidence. Although G1 withholds rejected content from the current consumption view, information derived from permitted access may still enter an agent's assessment and be passed to downstream agents. DEL therefore records these derived outputs together with their upstream dependencies, enabling G2 and G3 to enforce the corresponding authorization restrictions during downstream reuse and aggregation.

To evaluate the contribution of downstream enforcement, we conduct component ablation experiments using paraphrased attack payloads. The results are reported below, where lower ASR-any indicates better defense performance.

DatasetFull TrustGuardWithout G2Without G3
FairEval62.1%78.8%75.8%
LLMBar68.8%71.2%71.2%
MT-Bench71.1%71.1%69.7%

On FairEval, disabling G2 and G3 increases ASR-any by 16.7% and 13.7%, respectively. Sample-level logs further show that, after agent_1 accesses the memo and generates an opinion, Full TrustGuard tracks its source dependency and blocks its unauthorized downstream use, whereas disabling the relevant checks allows the opinion to propagate.

Direct comparison with G1-only
We further compare Full TrustGuard against G1-only in the three-judge, two-round setting using paraphrased attack payloads. Each dataset contains five candidate pairs. The results are shown below.

DatasetFull TrustGuardG1-onlyReduction in ASR-any
FairEval60%80%20 pp
LLMBar60%80%20 pp
MT-Bench80%100%20 pp

Overall, Full TrustGuard achieves lower ASR-any than G1-only across all three datasets. Such consistent improvements provide preliminary evidence that downstream dependency tracking offers additional protection beyond source-level authorization at G1.

AuthorsOpenReview ↗

Author Response (3/3): Clean Accuracy and Utility Trade-offs

3. Does TrustGuard preserve clean accuracy? (Weakness 2: Clean Accuracy)

Following your suggestion, we conducted an additional clean evaluation to directly compare TrustGuard with the undefended system. Using Qwen2.5-7B, we evaluate both methods on the same 222 candidate pairs without any attack payload. We report strict bidirectional accuracy, which counts a candidate pair as correct only when both the original and swapped candidate orders are evaluated correctly. The results are shown below.

MethodFairEval clean strict accuracyLLMBar clean strict accuracyMT-Bench clean strict accuracy
No Defense40.9%41.3%47.4%
TrustGuard40.9%41.3%47.4%

The results show that TrustGuard maintains the same measured clean accuracy as the undefended system across all three datasets in this evaluation. We will include these results in the revised paper and clarify the model and evaluation scope.

Additional utility ablation of downstream authorization checks. We further conduct an ablation study under a three-judge, two-round setting to examine the utility impact of downstream authorization checks. Unlike the clean evaluation above, this experiment compares Full TrustGuard with G1-only under two conditions: (1) no external memo, and (2) a benign memo containing information from an untrusted external source but no attack payload. We use the same strict bidirectional accuracy metric.

DatasetNo memo: FullNo memo: G1-onlyBenign memo: FullBenign memo: G1-only
FairEval39.4%39.4%37.9%31.8%
LLMBar37.5%37.5%31.2%37.5%
MT-Bench38.2%38.2%28.9%36.8%

The results indicate that without a memo, Full and G1-only produce identical sample-level outcomes, with no abstentions. With benign memos, the accuracy differences between Full and G1-only are +6.1%, −6.3%, and −7.9% across the three datasets, respectively. These results show that the additional authorization checks have a dataset-dependent impact on the use of benign external information, with accuracy decreases observed on two of the three datasets.

Accordingly, we will qualify our claim: TrustGuard maintains the measured clean accuracy in the tested no-attack configuration, while the additional downstream authorization checks may introduce some utility costs when processing benign external evidence. We will report both findings and their respective evaluation settings in the revised paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.