Skip to content

MIRAGE: Hiding LLM-based multi-agent topologies through phantom structure injection

Concealing LLM-Based Multi-Agent Topology via Phantom Structure Injection

arXiv· arXiv:2609.37567· Longzhu He, Zelang Wen, Xinfeng Li, Sen Su, XiaoFeng WangNanyang Technological University; Beijing University of Posts and Telecommunications; The Hong Kong Polytechnic UniversityIndexed by this site · 20 days ago· Originally publishedDefenses

Quick read

Defenses

Based on the paper PDF (AI-generated); refer to the original

The authors propose MIRAGE, a MAS topology hiding defense, and state that on three frameworks including G-Designer and four datasets it reduces the AUC/ACC/F1 of CIA topology inference while largely preserving task utility.

Threat modelUnder the strict black-box setting, the attacker submits self-chosen queries to the target MAS and observes the final responses, and can construct adversarial queries that induce responses to expose intermediate agent outputs in order to infer the topology; the attacker cannot access agent profiles, system prompts, intermediate messages, or the true topology G, and cannot modify internal system components.

Questions
The communication topology of an LLM multi-agent system carries design knowledge and IP, and the existing attack CIA can, under black-box conditions, infer the topology from the semantic dependencies of observable reasoning traces; the authors argue that the security dimension of topology confidentiality still lacks systematic defenses.
Methods
MIRAGE has three stages: synthesize a phantom topology whose structure differs sufficiently from the real topology, realize the phantom edges as plausible semantic dependencies and rewrite the traces of the hidden real edges, and still execute along the real topology while exposing externally only a view shaped according to the phantom topology.
Experiments and results
On the three frameworks G-Designer, AGP, and ARG-Designer and on MMLU, GSM8K, SVAMP, and HumanEval, MIRAGE reduces the AUC/ACC/F1 of CIA inference (e.g., on G-Designer+MMLU, AUC drops from 0.81 to 0.67), and the authors state that task utility changes only slightly.
Limitations and future work
The paper states that limitations and future directions are in Appendix E, but the provided text does not include the appendix content; the experiments cover only three topology optimization frameworks, four datasets, and 100 tasks per dataset, and the paper does not report the LLM used, the implementation model of the generator Mϕ, or the number of repetitions.
Experimental setupMMLU、GSM8K et al. · AUC、ACC et al.
Benchmarks
MMLUGSM8KSVAMPHumanEval
Metrics
AUCACCF1task accuracy
AI summaryResearchers propose MIRAGE, a topology-hiding framework that defends against attacks inferring the communication topology of LLM multi-agent systems from the semantic dependencies of reasoning traces in black-box settings. The framework has the system execute tasks according to the real topology and, through three stages—phantom topology synthesis, semantic edge realization, and protected execution—directs the semantic evidence visible to the adversary toward a constructed phantom topology. Experiments on 3 topology optimization frameworks and 4 benchmark datasets show that MIRAGE substantially weakens topology inference attacks while largely preserving task utility.

Researchers propose MIRAGE, a topology-hiding framework that defends against attacks inferring the communication topology of LLM multi-agent systems from the semantic dependencies of reasoning traces in black-box settings. The framework has the system execute tasks according to the real topology and, through three stages—phantom topology synthesis, semantic edge realization, and protected execution—directs the semantic evidence visible to the adversary toward a constructed phantom topology. Experiments on 3 topology optimization frameworks and 4 benchmark datasets show that MIRAGE substantially weakens topology inference attacks while largely preserving task utility.

Paper analysis

6 questions; open one to read the full answer

  1. What problem does this paper attempt to solve?

    How to protect the communication topology of an LLM multi-agent system from being reconstructed by black-box inference attacks without harming task performance.

    Core question: can the communication topology of an MAS be prevented from being reconstructed by black-box topology inference attacks without harming its task performance?

    • Problem setting: the communication topology of an LLM-based MAS determines the information flow among agents, and the authors argue that a carefully optimized topology encodes substantial design effort and constitutes valuable intellectual property (IP).
    • Why it matters: existing work CIA (Wu et al., 2026) shows that an attacker can reconstruct the topology using only black-box interaction by exploiting semantic dependencies in observable reasoning outputs; the authors argue that this exposes critical agents and communication paths, facilitating subsequent targeted attacks and bringing IP leakage risk.
    • Shortcomings of existing methods: the authors state that existing MAS defense research mainly targets security at the communication content level (e.g., malicious prompt propagation, communication manipulation), while systematic protection against topology inference is largely unexplored.
    • Core idea of the paper: treat
  2. What related research exists?

    Related research covers three categories: MAS and their topology optimization, attacks against MAS (including the topology inference attack CIA), and content-level MAS defenses.

    LLM-based MAS and topology optimization

    • ChatDev (Qian et al., 2024), CAMEL (Li et al., 2023), MetaGPT (Hong et al., 2024): early MAS use hand-designed communication structures, such as sequential workflows and role-based collaboration.
    • G-Designer (Zhang et al., 2025), AGP (Li et al., 2025), ARG-Designer (Li et al., 2026): automated topology optimization methods that build task-adaptive communication structures via graph generation or pruning; this paper uses these three frameworks to construct the protected target MAS.

    Attacks against MAS

    • Content-layer attacks: prompt-based attacks (Lee et al., 2025; Shahroz et al., 2025; Arif et al., 2026), communication attacks (He et al., 2025b; Yan et al., 2026), and task disruption (Amayuelas et al., 2024), which inject malicious instructions or manipulate inter-agent interactions.
    • CIA (Wu et al., 2026): under black-box conditions, reconstructs the MAS communication topology from semantic dependencies among agent outputs; the paper notes that its key insight is that directly connected agent pairs exhibit stronger semantic dependencies in observable reasoning outputs. This paper's defense targets this attack.

    Privacy and security defenses for MAS

    • BlindGuard (Miao et al., 2025), TRiSM (Raza et al., 2026), GUARDIAN (Zhou et al., 2025): protect agent collaboration via attack detection, trust management, or communication protection; the paper notes that these methods protect message content or system behavior, not the communication topology itself.
    • k-anonymity (Sweeney, 2002): this paper's topology anonymity set draws on its idea, hiding the true topology among multiple structurally feasible candidate topologies.

    Baselines and benchmarks

    • Baseline methods: the undefended original MAS (No Defense), and three methods commonly used for prompt injection defense—Instructional Prevention (Instruction), Delimiters, PPL Detection (Liu et al., 2024; Zhan et al., 2025), representing instruction-level prevention, input isolation, and detection-based defense, respectively.
    • Benchmarks/datasets: MMLU, GSM8K, SVAMP, HumanEval.

    The authors position this paper as a topology-confidentiality defense against black-box topology inference attacks such as CIA, filling a gap not covered by content-layer defenses.

  3. How does the paper address this problem?

    MIRAGE decouples the true topology used for task execution from the phantom topology visible to the adversary, realizing topology hiding in three stages.

    Overall idea: MIRAGE makes task execution strictly follow the real topology G, while shaping the semantic dependency evidence visible to the adversary toward a structurally different phantom topology G′, so that inference results deviate from G. The process is shown in Figure 2 and Algorithm 1.

    Problem Formalization

    • The MAS is represented as S = (P, G), where G is a DAG; the defense goal is to construct a protected system such that the topology inferred by the attacker differs maximally from G, while the utility loss is bounded: max Δtopo(Ĝ, G) s.t. U(S̃) ≥ U(S) − ϵ.
    • Here ∆topo measures the difference between two topologies, U is the task utility, and ϵ is the maximum allowed utility drop.

    Stage One: Phantom Topology Synthesis

    • Keep the original agent set A, and perturb only the communication edges; the candidate edge space Uπ is constrained by execution order to guarantee acyclicity, and the feasible topology set Ω(G) additionally requires that every agent has a path to the decision agent an.
    • Define the structural camouflage ratio (SCR) ρ(H, G) = 1 − |EH ∩ E|/|EH ∪ E|, which measures the structural deviation of a candidate topology from G; ρ = 1 means no shared edges.
    • Drawing on k-anonymity, propose the topology anonymity set: Kρ0(G) = {H ∈ Ω(G) | ρ(H, G) ≥ ρ0}, where the anonymity level Ktopo is the number of feasible topologies satisfying the minimum deviation threshold ρ0; construct this set via order-preserving edge rewiring and select G′, thereby obtaining the phantom edge set Ep = E′ \ E and the hidden real edge set Eh = E \ E′.

    Stage Two: Semantic Edge Realization

    • Phantom edge materialization: for each phantom edge (aj, ai), an LLM-based generator Mϕ generates semantic evidence conditioned on the source agent profile pj and output rj, the target profile pi, the query q, and the context ci: sphj→ i = Mϕ(pj, rj, pi, q, ci); generation follows three principles: role consistency, task relevance, and dependency plausibility.
    • Genuine edge obfuscation: for each agent ai's original trajectory ri, perform controlled paraphrasing (Bandel et al., 2022) to generate M candidates, and select according to $ \tilde{r}_i = \arg\min_{\hat{r}} [D_{sem}(\hat{r}, r_i) + \lambda \sum_{a_j \in P_i^h} S_{dep}(r_j, \hat{r})]$, suppressing the semantic dependency on the hidden real predecessors while preserving task semantics; Dsem measures semantic distortion, Sdep measures semantic dependency, and λ balances the two.
    • Both operations act only on the adversary-facing context and do not change the actual communication topology.

    Stage Three: Protected MAS Execution

    • Execution view: each agent still receives predecessor outputs and executes according to G; the phantom topology does not participate in internal communication.
    • Exposure view: before external exposure, the exposure view is assembled using the obfuscated trajectories and phantom evidence, so that the observable dependency structure deviates from G and tends toward G′.

    Evaluation Method

    • Topology protection is measured by the AUC, ACC, and F1 of CIA inference (AUC close to 0.5 and lower ACC/F1 indicate stronger protection); task utility is measured by the accuracy on each benchmark; the parameters ρ0, M, and λ are selected from {0.2,…,1.0}, {1,…,9}, and {0.1,…,0.9}, respectively.
  4. What experiments did the paper conduct?

    Evaluated on three topology optimization frameworks and four datasets: MIRAGE reduces the AUC/ACC/F1 of CIA inference, and the authors state that task utility changes only slightly.

    Experimental setup

    • Protected systems: target MAS constructed by three topology optimization frameworks—G-Designer, AGP, ARG-Designer; the paper does not report the LLM driving these MAS.
    • Datasets: MMLU (general reasoning), GSM8K and SVAMP (mathematical reasoning), HumanEval (code generation); 100 tasks are sampled from each dataset (Section 5.1).
    • Attack: CIA topology inference attack (Wu et al., 2026).
    • Baselines: No Defense; three prompt injection defenses, Instruction, Delimiters, PPL Detection.
    • Metrics: AUC, ACC, F1 of topology inference (computed over candidate communication edges); task utility is accuracy.
    • Topology scale (Table 2): average number of agents 3.05–7.00, average number of edges 3.10–11.54; ARG-Designer's topologies are smaller than those of the other two frameworks.

    Main results

    Table 1 gives the topology inference comparison of MIRAGE and two baseline defenses on G-Designer (lower AUC/ACC/F1 means stronger protection):

    Swipe horizontally to read the table

    MethodMMLU AUC/ACC/F1GSM8K AUC/ACC/F1SVAMP AUC/ACC/F1HumanEval AUC/ACC/F1
    Instruction0.79/0.59/0.620.81/0.73/0.640.72/0.74/0.690.78/0.62/0.59
    Delimiters0.76/0.61/0.640.80/0.76/0.670.70/0.68/0.710.76/0.65/0.56
    MIRAGE (Ours)0.67/0.50/0.490.64/0.51/0.420.64/0.48/0.390.68/0.44/0.52

    (According to Table 1)

    • Cross-framework results: Figures 3–5 show that MIRAGE reduces CIA's AUC/ACC/F1 on all three frameworks; for example, on G-Designer+MMLU AUC drops from 0.81 to 0.67 (Figure 3), on AGP+MMLU from 0.89 to 0.79 (Figure 4), and on ARG-Designer+GSM8K from 0.98 to 0.78 (Figure 5).
    • Authors' interpretation: the authors state that MIRAGE consistently lowers inference performance across all three topology optimization strategies and four datasets, and that Table 2 shows large differences in topology complexity across settings, which the authors take as evidence of the method's robustness.
    • Exception: in most settings MIRAGE's AUC remains above the random level of 0.5 (e.g., 0.64–0.68 across datasets on G-Designer), i.e., inference is not fully reduced to random guessing.

    PPL detection baseline

    • What was measured: the ability of PPL-based detection to distinguish topology inference queries from benign inputs; Figure 6 gives ROC curves on the four datasets.
    • Results: detection AUC is 0.51 on MMLU, 0.57 on GSM8K, 0.60 on SVAMP, and 0.59 on HumanEval, close to the random-guess reference line.
    • Authors' interpretation: the authors present this result as one of the existing defense evaluations compared with MIRAGE; the main text does not further explain why it is close to random.

    Task utility (RQ2)

    • What was measured: Figure 7 compares task accuracy before and after defense (Original vs Ours) on three frameworks and four datasets.
    • Results: the bar values of Figure 7 are unreadable in the converted text; the main text states that MIRAGE largely preserves original task utility across all evaluated settings, with only minor performance variations, without giving specific numbers.

    Ablation and parameter analysis (RQ3, G-Designer + MMLU)

    • Component ablation: Figure 8(a) shows that removing phantom edge materialization (w/o PEM) or genuine edge obfuscation (w/o GEO) both raise AUC, ACC, and F1, and the authors state that both components are necessary for topology hiding (specific values are unreadable in the converted text).
    • Parameter ρ0: Figure 8(b) shows a trade-off for ρ0 between topology hiding and task utility.
    • Parameter M: Figure 8(c) shows that the benefit of increasing the number of rewrite candidates M saturates after M = 5.
    • Parameter λ: Figure 8(d) shows that a moderate λ balances dependency suppression and semantic preservation better.

    Case study (RQ4)

    • Figure 9 visualizes the topologies inferred by CIA under the three frameworks: without defense the inferred structure closely resembles the ground truth, while with MIRAGE the inferred topology deviates from the ground truth, with multiple true edges hidden or replaced by phantom relations (per the authors).
  5. What are some points that can be further explored?

    The paper states that limitations and future directions are in Appendix E, but the provided text does not include the appendix; the experiments cover three frameworks, four datasets, and 100 tasks per set.

    Limitations and future directions noted by the authors

    • The paper states in the Conclusion (Section 7) that discussion of limitations and future directions is in Appendix E, but the provided text does not include the appendix content, so specific items cannot be listed.

    Experimental coverage

    • Protected systems: MAS constructed by three topology optimization frameworks, G-Designer, AGP, and ARG-Designer (Section 5.1).
    • Datasets: four datasets, MMLU, GSM8K, SVAMP, and HumanEval, with 100 tasks sampled from each (Section 5.1).
    • Attack: only one topology inference attack, CIA, is evaluated (Sections 2, 5).
    • Baseline defenses: No Defense, Instruction, Delimiters, PPL Detection (Section 5.1).
    • Not reported by the paper: the LLM driving the MAS, the model used by the generator Mϕ, the number of experimental repetitions, and statistical significance tests; ablation and parameter analyses are conducted only on G-Designer + MMLU (Section 5.2).
  6. Summarize the main content of the paper

    MIRAGE decouples the semantic evidence exposed externally by the MAS from the real execution topology through phantom topology injection, thereby reducing the effectiveness of black-box topology inference.

    A defense paper on the topology confidentiality of LLM multi-agent systems.

    • Problem: topology inference attacks such as CIA can, under black-box conditions, reconstruct the communication topology from semantic dependencies in the observable outputs of a MAS; the authors consider this to pose two risks, exposure of structural vulnerabilities and IP leakage, while existing MAS defenses mostly target communication content.
    • Method: MIRAGE has three stages—synthesize a phantom topology G′ that structurally deviates from the true topology G by at least ρ0 under order-preserving and connectivity constraints; use an LLM generator to realize phantom edges as credible semantic dependencies and, via controlled rewriting, suppress the dependency traces of hidden true edges; task execution still follows G, and only a view shaped according to G′ is exposed externally.
    • Main results: on three frameworks, G-Designer, AGP, and ARG-Designer, and on MMLU, GSM8K, SVAMP, and HumanEval (100 tasks per set), MIRAGE reduces the AUC/ACC/F1 of CIA inference; on G-Designer, MIRAGE's metrics are all lower than the two baselines Instruction and Delimiters (Table 1), e.g., AUC 0.67 vs 0.79/0.76 on MMLU.
    • Other findings: PPL detection's AUC for distinguishing inference queries is 0.51–0.60, close to random (Figure 6); the authors state that task utility changes only slightly (Figure 7, specific numbers unreadable); ablations show that both components, phantom edge materialization and genuine edge obfuscation, are indispensable (Figure 8(a)).
    • Authors' conclusion: the authors state that MIRAGE reduces topology inference effectiveness across multiple topology optimization methods and datasets while largely preserving task utility, and is a practical approach to protecting the confidential topology of MAS; limitations and future directions are in Appendix E (the provided text does not include that appendix).

Related papers

Show 3 more papers
Read originalarxiv.org