acceptodds
Under review as a conference paper at ICLR 2027

Federated Risk Learning for Adaptive Post-Quantum IoT Authentication

Abstract

Smart-home Internet of Things (IoT) security requires behavioral adaptation without exposing household traffic or weakening authentication during post-quantum migration. Existing work advances private learning, robust aggregation, and cryptographic authentication, but leaves open how uncertain device behavior should guide a certified key lifecycle. This work presents FALCON-SDIoT, a software-defined networking architecture that connects federated risk classification to authenticated rekeying. Manufacturer profiles guide initial classification, clipped and noised edge updates support collaborative learning, and temporal trust determines asymmetric hardening and relaxation. Certified key establishment and nonce-bound proof-of-possession bind these decisions to device credentials. This work integrates federated risk assignment, credential binding, and coordinated class and key transitions under explicit privacy and security assumptions. Evaluation covers a controlled synthetic benchmark, public-traffic temporal and complete-device holdouts, repeated-seed federated tests, and an adaptive-policy sensitivity study. Results support manufacturer-informed risk classification, characterize performance under adversarial updates, and demonstrate the trade-off between detection delay and false alarms. Public-traffic tests assess temporal classification and unseen-device detection. These findings support the component interfaces under the evaluated settings; cryptographic costs remain accounting estimates, and the recurring proof is assessed within a classical random-oracle scope. The architecture supports studying behavior-driven authentication while distinguishing predictive confidence from cryptographic assurance.

Primary area:
infrastructure, software libraries, hardware, systems, etc.
Keywords:
federated learning, iot security, adaptive authentication, privacy-preserving machine learning, byzantine-robust aggregation, post-quantum cryptography, device classification, anomaly detection

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_1N7qOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FALLCON-SDIoT, combining federated risk classification, post-quantum credentials, anomaly-based trust adaptation, and risk-triggered rekeying for smart-home IoT. The problem is relevant and the paper is clearly written, but the current evidence does not validate the integrated system.

Critical strengths
  1. The paper addresses an important problem, clearly distinguishes behavioral confidence from cryptographic assurance, states several assumptions and limitations explicitly.

  2. The writing and logic is clear and easy to follow.

Critical weaknesses
  1. The post-quantum security claim is only partially substantiated. Although the protocol employs standardized ML-KEM-768 and ML-DSA-65, its recurring proof of possession is left parameterized and Theorem 1 assumes, rather than establishes, the knowledge soundness of the underlying Fiat–Shamir construction.

  2. Federated learning appears to determine only the initial risk class, while subsequent class transitions are driven by a separate local Isolation Forest. It is therefore unclear how federated risk learning actually guides the adaptive authentication lifecycle after enrollment.

  3. The contribution to federated learning is limited. The method mainly combines existing techniques—profile initialization, clipping, Gaussian noise, and Multi-Krum—without proposing a new FL algorithm or analysis. Its contribution is better characterized as system integration.

  4. The federated evaluation lacks necessary baselines and ablations. There are no comparisons with FedAvg, coordinate median, trimmed mean, standard Krum, adaptive clipping, or learnable aggregation.

  5. Theorem 1 assumes an already knowledge-sound Fiat–Shamir proof and essentially restates the resulting possession guarantee. No concrete proof protocol is instantiated or benchmarked. Moreover, the paper lists ProVerif targets without clearly reporting successful verification.

Questions for authors

See weaknesses

1 comment
AuthorsOpenReview ↗

Official response to Reviewer 1N7q

We thank the reviewer for the careful assessment and for recognizing the distinction between behavioral confidence and cryptographic assurance. The concerns identify several areas where the contributions and evidence require clearer boundaries.

1. Post-quantum proof and Theorem 1.

We agree that Theorem 1 is conditional rather than a new proof of lattice-based knowledge soundness. It assumes a complete, knowledge-sound Fiat–Shamir proof for the bounded-witness relation in Equation (13), together with unforgeable certification and fresh, single-use nonces. Under these assumptions, it establishes the implication from accepted authentication to extractable possession of a witness for the certified public key.

The theorem does not establish the knowledge soundness of a concrete proof instantiation, quantum random-oracle security, or native post-quantum proof performance. ML-KEM-768 and ML-DSA-65 provide standardized cryptographic components, but their use does not remove these limitations. We will state this distinction more prominently.

2. Role of federated learning after enrollment.

The reviewer is correct. Equation (5) uses the manufacturer-initialized federated classifier to assign the initial risk class. Subsequent transitions are governed by local Isolation Forest anomaly evidence and the temporal trust rules in Equations (15)–(16).

The connection is therefore through a shared device-policy state: federated learning establishes the initial class, local monitoring adjusts that class, and Equation (4) coordinates class changes with authenticated key installation. The current design does not continuously feed federated predictions into every subsequent transition. We will clarify this two-stage architecture rather than imply continuous federated control of authentication.

3. Contribution beyond existing federated methods.

We agree that clipping, Gaussian noise, Multi-Krum, and Isolation Forest are established techniques. The methodological contribution is the coordination of manufacturer-informed risk assignment, credential verification independent of predictions, asymmetric trust adaptation, and authenticated class/key commitment. This is a system-level contribution, not a claim of a new federated learning algorithm.

4. Missing baselines and ablations.

The current experiments provide a manufacturer-profile initialization ablation: mean synthetic A/B/C risk accuracy improves from 78.05% to 91.58%. They also characterize two non-adaptive Byzantine-update settings and evaluate 81 temporal-policy configurations.

However, the paper does not include matched comparisons against FedAvg, coordinate median, trimmed mean, standard Krum, adaptive clipping, or learnable aggregation. We agree that these comparisons are needed to isolate the contribution of the aggregation configuration. In particular, changing participant counts across the Byzantine settings prevents a direct robustness ranking. We will clarify these limitations and distinguish the available ablations from the missing comparisons.

5. ProVerif verification status.

The reviewer correctly observes that Section 4.3 lists six symbolic verification targets without reporting successful verification results. These targets must not be interpreted as established protocol guarantees. We will explicitly separate the proposed verification queries, the conditional theorem, and properties requiring concrete verification.

Overall, FALCON-SDIoT should be evaluated as a proposed architecture linking risk learning to credential-bound authentication transitions, with component-level evidence and explicitly limited security guarantees. We appreciate these comments and will revise the manuscript to make this contribution and its limitations more precise.

Reviewer_L76fOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FALCON-SDIoT, a federated risk learning framework for adaptive post-quantum IoT authentication. It integrates manufacturer-informed risk classification, privacy-aware federated updates, credential verification, and trust-based key management. The evaluation includes synthetic classification experiments, public traffic analysis, and simulated authentication policies.

Critical strengths
  1. Relevant problem formulation. The paper addresses an important practical challenge by connecting federated risk assessment with adaptive authentication in IoT environments. Separating behavioral predictions from cryptographic assurance is a reasonable design choice.
  2. Structured system design. The proposed framework provides a coherent architecture for risk classification, credential verification, and key lifecycle management. The authors also acknowledge several limitations of their security assumptions and evaluation scope.
Critical weaknesses
  1. Limited methodological novelty. The framework primarily integrates established techniques, including Multi-Krum, Gaussian perturbation, Isolation Forest, and standardized post-quantum cryptography. The paper does not sufficiently demonstrate a novel learning mechanism or a substantial theoretical contribution beyond their integration. This limits its significance for ICLR.
  2. Insufficient end-to-end validation. The central federated risk classification experiments rely on synthetic data, while the public traffic evaluation focuses on device identification and unseen-device detection rather than authentication risk prediction. Moreover, the post-quantum proof is not concretely instantiated, and the reported latency and energy results are analytical estimates. Consequently, the practical effectiveness of the complete framework remains insufficiently demonstrated.
Questions for authors

Q1. What is the primary methodological contribution beyond combining existing federated learning, anomaly detection, and authentication techniques? Can the authors provide a clearer comparison with closely related integrated frameworks?

Q2. Can the authors demonstrate the complete risk-to-authentication pipeline using real traffic and an executable cryptographic implementation? How would the conclusions change when accounting for actual proof generation, communication overhead, and rekeying failures?

2 comments
AuthorsOpenReview ↗

Official response to Reviewer L76f

We thank the reviewer for the constructive assessment and for recognizing the relevance of connecting federated risk assessment with adaptive authentication. We agree that the methodological contribution and experimental scope need to be distinguished more clearly.

Q1. What is the methodological contribution beyond integrating established techniques?

The contribution is not a new federated optimizer, anomaly detector, or post-quantum primitive. Rather, FALCON-SDIoT addresses how uncertain behavioral evidence can guide authentication decisions without weakening credential assurance or creating inconsistent security states.

Three interfaces define this contribution:

First, risk-to-policy assignment (Section 3.3): Manufacturer profiles initialize federated A/B/C risk classification while raw traffic remains at gateways. On the synthetic benchmark, this initialization improves mean clean risk accuracy by 13.53 percentage points over the no-profile configuration.

Second, independent credential assurance (Section 3.4): Device authentication requires certified secret possession regardless of predicted risk. Behavioral confidence cannot substitute for cryptographic verification.

Third, consistent policy transitions (Section 3.5): Temporal trust controls asymmetric hardening and relaxation, while Equation (4) commits the candidate risk class and transport key only after authenticated installation. Uncertain installation suspends service pending recovery.

The distinction from closely related work is at this interface level. CALIS (2026) studies controller-driven cipher selection and key rotation using classical key establishment, while Rehman et al. (2026) combine quantum-resilient federated anomaly detection with trust adaptation. FALCON-SDIoT specifically connects manufacturer-informed risk assignment, certified possession, and class/key commitment. We will make this comparison more explicit and avoid presenting established components as individually novel.

Q2. Can the complete pipeline be demonstrated using real traffic and executable cryptography?

We agree that the current evaluation does not establish complete end-to-end deployment performance. The evidence covers distinct components: synthetic federated risk classification, real-traffic device identification and unseen-device detection, conditional credential analysis, and scripted temporal policy evaluation. Public-traffic experiments do not directly validate A/B/C authentication-risk prediction.

The recurring authentication costs of 10.69 ms and 123.41 mJ are analytical estimates. They do not include native proof generation, onboarding, ML-KEM setup, training, or long-term monitoring. Real proof execution, packet fragmentation, retries, and rekeying failures could increase latency and energy. Equation (4) specifies the intended fail-closed behavior, but distributed failure recovery has not been experimentally validated. Therefore, the estimated cost advantage should not be interpreted as measured end-to-end superiority.

We will revise the presentation to distinguish implemented evaluations, analytical estimates, conditional security guarantees, and deployment requirements. This preserves the paper's central contribution as a risk-guided authentication architecture while avoiding claims beyond the available evidence.

We appreciate the reviewer's suggestions and believe these clarifications will make both the novelty and the current validation boundaries substantially easier to assess.

Thank you for the authors' responses and clarifications. While I appreciate the effort, my main concerns regarding methodological novelty and end-to-end validation remain. Therefore, I maintain my original assessment.

Reviewer_wAnDOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper sets out to use device behaviour profiles with authentication, to manage credential life cycle. Some device behaviour and perceived risk would trigger credential management. The proposed approach, FALCON-SDIoT, is a software-defined networking architecture that connects the federated risk classification to authenticated rekeying. Manufacturers’ profiles are used for device classification, and some risk profiles. Then standard key establishment protocols are used to establish credentials. This approach supports behaviour-driven re/authentication. The work is evaluated from the perspective of device ID/behaviour classification and also since IoT communication time overhead and energy.

Critical strengths

The idea of risk guided credential management is interesting. Periodic authenticatin has overhead, challenging a device only if suspect is a logical and elegant idea.

Critical weaknesses

For a security paper, there is not a lot of discussion on what the actual risk behaviour is. "explicit privacy and security assumptions" are mentioned but not explained. The risk is treated as quite a high level concept. No security model, or specific actions. Various attacks are mentioned throughout, but it is not systematic. It shoud state what are the main security issues you want to detect in the behaviour, what are the things the attacker will do to hide this, and what is the performance. Especially also what is the performance in comparision to related work. The only clear comparison provided here is on latency and energy. This is where the contribution is - device identification/classification is around a long time, related to this device risk classification through IDS, etc.

The paper is really difficult to follow - I have read many security and crypto papers over decades. Usually the maths/proofs are dense and difficult to follow, but here it is the actual writing style. Simply put, it seems any normal word or statement is somehow condensed and upgraded to the most complex synonym/equivalent, and then the sentence do not combine to make sense.

Example: "Manufacturer profiles guide initial classification, clipped and noised edge updates support collaborative learning, and temporal trust determines asymmetric hardening and relaxation. Certified key establishment and nonce-bound proof-of-possession bind these decisions to device credentials." Here are two sentences from the abstract. Both contain significant amount of information - difficult to read as not clear what the "decisions" are? Are these decisions the classification decision?

Also there are many sentences like these. Yes, these are grammatically correct, but it is essentially like shorthand notes: Forward secrecy concerns signing-key compromise after ephemeral-secret erasure; active-key confidentiality assumes trusted gateways and controllers. The proof is parameterized; native verification costs remain unmeasured.

Questions for authors

What is the nature of the comparative works in Figure 4? Are they doing similar risk-based authentication management or are these works dealing mostly with key establishement/authentication. Where is the major gain here? It says this is mostly only the crypto and comms - is the crypto protocol the one in 3.4?

Are there any comparative works on the device identification? Are there benchmarks for other results?

2 comments
AuthorsOpenReview ↗

Official response to Reviewer wAnD

We thank the reviewer for the detailed assessment and for recognizing the value of risk-guided credential management. We agree that the manuscript does not explain the security objectives and their connection to device behavior clearly enough. We address the main concerns below.

1. Security assumptions and behavioral risk. Section 3.1 considers two adversarial settings: an active network attacker capable of injecting, modifying, delaying, or replaying messages, and Byzantine gateways submitting malicious federated updates. The controller and serving gateway are trusted for credential assurance, and their communication channel is authenticated and confidential.

Section 3.5 monitors inter-arrival times, traffic volume, protocol proportions, destination-port entropy, connection attempts, and TLS usage. These features provide evidence of abnormal behavior, but do not establish that a specific attack has occurred. The current evaluation covers scripted anomalous behavior and non-adaptive Gaussian update attacks; it does not establish detection performance against specific malware or evasive traffic attacks. We will make this distinction explicit.

2. Concrete authentication actions. The decision process is as follows:

  • Manufacturer-informed federated learning assigns the initial A/B/C risk class (Eq. 5).
  • Local anomaly evidence updates temporal trust (Eq. 15).
  • Low trust triggers one-step hardening, while sustained high trust permits gradual relaxation (Eq. 16).
  • A class change or renewal request triggers a new transport key (Eq. 12), and authenticated installation commits the class and key together (Eq. 4).
  • Failed authentication denies service, while uncertain key installation suspends service until recovery.

Importantly, recurring proof of possession remains independent of behavioral risk. Risk determines policy transitions rather than replacing credential verification.

3. Figure 4 and comparative performance. The six schemes in Figure 4 primarily concern authentication or key establishment, not equivalent risk-adaptive credential management. Figure 4 estimates recurring authentication costs using literature-normalized inputs. The reported 10.69 ms and 123.41 mJ are analytical estimates, not measurements of an implemented end-to-end system. They exclude enrollment, ML-KEM setup, federated training, and continuous monitoring. Section 3.4 defines the credential lifecycle, but the recurring proof has not been instantiated or natively benchmarked. Accordingly, Figure 4 cannot establish an end-to-end performance advantage over risk-aware systems.

4. Classification benchmarks. Table 1 compares Stacking (90.86%) and Naive Bayes (90.80%) for 27-device identification. For the separate synthetic A/B/C risk task, manufacturer-profile initialization improves mean accuracy from 78.05% to 91.58%. Public-traffic evaluation achieves 96.47% known-device identification accuracy and 0.9564 unseen-device AUROC. We acknowledge that these public-traffic results do not constitute a matched comparison of real-world authentication-risk prediction.

5. Readability. We agree that several passages are unnecessarily compressed. We will simplify the abstract and technical explanations, introduce a clearer threat-to-observation-to-action description, and distinguish the evaluated evidence from proposed security mechanisms.

The central contribution is the connection between behavioral risk decisions, independently verified credentials, and consistent class/key transitions. We appreciate the reviewer's comments, which identify important opportunities to communicate this contribution more clearly.

I appreciate the time takes by the authors to respond. I now understand some parts better but my overal opinion of the paper has not changed. As it is currently written, it is is goig to take significant work to update it to a state where it is acceptable and this is not going to be achievable for this event.

I suggest you consider the review comments and your response, and significant redesign this paper to have a very clear novel idea and a clear, simple threat model with logical attacks to test the system and allowing for comparison to other work. As mentioned, IDS area has so much work - not being able to identify systems that are comparable then starts to look more like solving a problem that does not need to be solved. If your idea can cover the same scenarios as prior work, improve on thes and add something new then we have a reasonable contribution.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.