Skip to content
Crypto.world

Research

Adaptive Trust Consensus for Blockchain IoT: Comparing RL, DRL, and MARL Against Naive, Collusive, Adaptive, Byzantine, and Sleeper Attacks

Trust-based consensus lets a blockchain IoT network pick its consensus committee from the nodes that have behaved well, which saves computation on weak devices — but it also turns trust scores…

Adaptive Trust Consensus for Blockchain IoT: Comparing RL, DRL, and MARL Against Naive, Collusive, Adaptive, Byzantine, and Sleeper Attacks
arXiv
2512.22860
Published
2025-12-28
Authors
Soham Padia, Dhananjay Vaidya, Ramchandra Mangrulkar

AI summary

Overview

Research area: Security and privacy of blockchain-enabled Internet of Things (IoT) networks — specifically trust-based consensus, adversarial robustness, and reinforcement learning for defense (cs.CR, with intersections in distributed systems and applied machine learning).

Technical level: Advanced. The paper assumes familiarity with reinforcement learning (tabular Q-learning, Dueling Double DQN, multi-agent RL), Bayesian trust models, Byzantine fault tolerance, and cryptographic access control (Fully Homomorphic Encryption and Attribute-Based Access Control).

Scope in one sentence: The paper builds a trust-based delegated consensus framework for permissioned blockchain IoT networks, secures its access-control layer with FHE-backed ABAC, and systematically compares three reinforcement learning defenses (RL, DRL, MARL) against five simulated attack families on a 16-node network.

What This Paper Is About

Trust-based consensus lets a blockchain IoT network pick its consensus committee from the nodes that have behaved well, which saves computation on weak devices — but it also turns trust scores themselves into an attack surface. Adversaries can inflate each other's scores, learn and evade the defense, send conflicting messages, or behave honestly for a long stretch and then strike. The paper's goal is to find out which of three learning-based defenses — simple tabular Q-learning, deep RL, or multi-agent RL — actually holds up against each of these five threat styles under identical conditions.

Key Contributions

  1. A five-family attack taxonomy for trust-based consensus. The authors formalize and implement the Naive Malicious Attack (NMA), Collusive Rumor Attack (CRA), Adaptive Adversarial Attack (AAA), Byzantine Fault Injection (BFI), and Time-Delayed Poisoning (TDP), covering naive, coordinated, learning-based, Byzantine, and temporal ("sleeper agent") threat models in one testbed.

  2. A controlled three-way comparison of learning paradigms. The paper states it provides the first systematic comparison of tabular RL, DRL (Dueling Double DQN / D3QN), and MARL for malicious-node detection under identical simulation conditions across all five attacks.

  3. A privacy-preserving trust framework. It couples an FHE-secured ABAC layer with the trust-based delegated consensus, so access-control policy evaluation happens on encrypted attributes and does not leak which attributes triggered an accept or reject decision — which the authors argue matters specifically against adaptive adversaries probing the defense boundary.

  4. An empirical security analysis identifying where coordination helps and where it does not. MARL beats the single-agent methods under collusive attack, but all three agents collapse under the delayed sleeper attack, which the authors flag as a critical open vulnerability.

Main Findings

  • MARL wins under collusion, but by a contested margin. The abstract reports MARL reaching F1 = 0.85 against the Collusive Rumor Attack versus DRL 0.68 and RL 0.50. The introduction reports the same comparison as MARL 0.85, DRL 0.62, RL 0.50 — the DRL figure differs between the two passages and the paper does not reconcile it in the provided text.

  • Adaptive attacks defeat the simplest agent entirely. Against the Adaptive Adversarial Attack, both DRL and MARL reach F1 = 1.00, while tabular RL stalls at F1 = 0.50.

  • Byzantine behavior is fully handled by every agent. All three approaches achieve F1 = 1.00 against Byzantine Fault Injection, so this attack family does not discriminate between the paradigms.

  • Naive attacks are not where the paradigms separate meaningfully — the headline differentiation in the abstract comes from collusive and adaptive attacks, not from the uncoordinated baseline threat.

  • Sleeper attacks are catastrophic for every agent. After the Time-Delayed Poisoning agents activate, F1 scores fall to 0.11–0.16 across all three approaches, the worst result reported anywhere in the paper.

  • Trust is deliberately asymmetric. Confirmed malicious behavior both increments the negative Beta parameter and multiplicatively decays the positive one, encoding the principle that trust should be harder to earn than to lose.

  • Delegation ratio is the single learned control knob. Rather than tuning detection thresholds, the agents act on a three-way discrete action space (multiply the delegation ratio by 0.9, hold at 1.0, or multiply by 1.1), with the ratio clipped to [0.1, 1.0].

  • The reward function is security-weighted. Detection quality carries weight 0.7 (F1 scaled by 100) and operational metrics weight 0.3, while each false negative is penalized at weight 3.0, reflecting the asymmetric cost of missing a malicious node.

Methodology in Plain English

The authors simulate a small permissioned blockchain IoT network of 16 nodes on which about 30% (roughly 5 nodes) are malicious, sitting at the Byzantine one-third boundary. Each node carries a trust score modeled as a Beta distribution — a running "benefit of the doubt" estimate that tracks both how trustworthy a node looks and how much evidence exists behind that estimate. Good behavior raises the positive counter, misbehavior raises the negative one, and confirmed malice also shrinks the positive counter. Consensus work is handed to a committee of the most trusted nodes, chosen by Thompson Sampling (drawing a random sample from each node's trust distribution and taking the top ones), which keeps uncertain nodes in play rather than letting early bad luck permanently exile an honest device.

Access control is handled by attribute-based rules (device role, permissions, clearance, live trust score), but the rules are evaluated on encrypted attributes via Fully Homomorphic Encryption so an eavesdropper cannot infer the decision boundary by probing.

The defense itself is framed as a Markov decision process. The agent sees a 16-dimensional state vector — seven summary statistics of the trust distribution (mean, variance, skewness, median, range, interquartile range, coefficient of variation) plus operational signals such as normalized verified transactions, chain length, honest-to-malicious ratio, fractions of low- and high-trust nodes, delegation efficiency, throughput, recent block creation rate, and a collusion-detection score. It then picks one of three delegation-ratio adjustments each step and receives a reward blending F1, operational performance, false-negative penalties, and an extra collusion penalty that fires when a collusion score (the reciprocal of the gap between mean honest trust and mean malicious trust) exceeds 2.0, capped at 20.

The three agents differ in representation. Tabular Q-learning discretizes the state and looks up a table; DRL uses a Dueling Double DQN that splits value and advantage estimation and uses separate online and target networks with experience replay; MARL runs N agents (one per node) that share neural network parameters but keep separate experience buffers and a shared global reward. Each attack family is then implemented as a distinct adversary policy — from random noise, through mutual trust boosting plus coordinated penalty of high-trust honest nodes, through an ε-greedy adversary rotating among five strategies (gradient exploitation, slow poisoning, strategic cooperation, mimicry, temporal coordination), through equivocation with virtual Sybil identities and eclipse attacks, to sleeper agents that behave honestly for 25 episodes before activating.

Why This Matters

Research impact. Trust-based consensus is usually evaluated against one or two attack types, typically a naive or collusive adversary. By putting five threat models and three learning paradigms in one controlled testbed, the paper produces a comparison table that other trust-consensus researchers can reuse, and it isolates a specific negative result — that trust-building sleeper agents break every tested defense — that redirects attention from detection accuracy toward long-horizon trust memory.

Real-world applications:

  • Industrial and smart-factory IoT, where device fleets are heterogeneous in compute power and a compromised sensor can persist undetected for long periods before acting.
  • Smart-home and building access control, where attribute-based rules must be enforced without exposing occupancy or permission data to the enforcement infrastructure.
  • Supply-chain and asset-tracking networks, where participants have commercial incentives to collude and mutually vouch for one another.
  • Critical-infrastructure sensor networks and smart grids, where Byzantine equivocation and network partitioning are realistic failure modes and missing a malicious node is far costlier than a false alarm.

Industry relevance. The paper's framing — a learned controller that tunes how many nodes form the consensus committee — is attractive to operators because it avoids hand-tuned static thresholds and trades a small amount of compute for measurable security. The FHE-backed ABAC component speaks directly to regulated deployments where policy logic itself is sensitive. The sleeper-agent result is the commercially important caveat: any vendor claiming trust-score-based security should expect that a patient adversary will pass the observation window.

Future Directions

  • Countermeasures for Time-Delayed Poisoning. The paper explicitly frames TDP as a critical vulnerability requiring dedicated defenses; the obvious research thread is long-horizon trust memory (retaining historical behavior patterns rather than weighting recent conduct) and delayed-credit or decay-on-activation scoring schemes.

  • Scaling beyond 16 nodes. The simulation uses a small-to-medium network with 5 malicious nodes; whether MARL's collusion advantage (and its parameter-sharing design) survives larger, more heterogeneous topologies is untested in the reported results.

  • Measuring the FHE overhead directly. The framework integrates FHE-secured ABAC, but the results reported in the provided content focus on detection F1 under attack; the latency and throughput cost of homomorphic policy evaluation on resource-constrained devices is an open question here.

  • Reconciling and hardening evaluation. Reproducing the collusive-attack DRL figure (reported as both 0.68 and 0.62 in different sections) and extending the comparison to stronger adaptive adversaries — for example, adversaries that also learn across episodes rather than selecting among five fixed strategies — would sharpen the conclusions.

Target Audience

Graduate students and researchers working on blockchain-IoT security, trust and reputation systems, or adversarial multi-agent reinforcement learning will get the most from this paper, particularly those looking for a ready-made five-attack benchmark or for evidence on when multi-agent coordination actually pays off. Security architects and protocol designers evaluating trust-based consensus for resource-constrained deployments will find the failure modes and the sleeper-agent warning practically useful. Readers without a background in reinforcement learning or Bayesian trust modeling will need to work through Sections 3 and 4 carefully, since the comparison hinges on architectural details (state representation, reward weighting, parameter sharing) rather than on a single headline metric.

Authors’ abstract

Securing blockchain-enabled IoT networks against sophisticated adversarial attacks remains a critical challenge. This paper presents a trust-based delegated consensus framework integrating Fully Homomorphic Encryption (FHE) with Attribute-Based Access Control (ABAC) for privacy-preserving policy evaluation, combined with learning-based defense mechanisms. We systematically compare three reinforcement learning approaches -- tabular Q-learning (RL), Deep RL with Dueling Double…

The full abstract on arXiv
Read the original paper