Agentic SABRE: An Uncertainty-Aware Neuro-Symbolic Multi-Agent Framework for Adaptive Ransomware Detection¶

🕒 Published (v1): 2026-07-05 13:08 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

TL;DR¶

Agentic SABRE is a neuro-symbolic, uncertainty-aware multi-agent framework for ransomware detection that fuses a semantic agent (PE embeddings, \(\mathbb{R}^{384}\)) and a behavioural agent (sliding-window I/O telemetry, \(\mathbb{R}^{16}\)) via a decision-layer orchestrator. The orchestrator applies a formal triage policy over fused risk score \(\hat{p}\) and aggregated epistemic uncertainty \(\sigma_{\max}\) to route samples to AUTO-CONTAIN, ESCALATE, or ALLOW. On RDset and RanSMAP, it achieves AUC = 1.0 on saturated semantic data and a 4.9% relative reduction in false escalations at equal recall.

Problem¶

Static monolithic classifiers fail under concept drift, evasion, and behavioural polymorphism: they produce point predictions with no epistemic uncertainty, which makes it impossible to calibrate the boundary between autonomous containment (where false positives are costly) and human escalation (where delays are costly). Existing uncertainty-aware systems treat uncertainty as a diagnostic output rather than as an operational control signal driving executable decisions.

Method¶

SABRE operates as a five-stage pipeline:

Hybrid feature extraction: static PE file embeddings (semantic) and time-windowed I/O telemetry statistics (behavioural), enriched with ATT&CK-inspired symbolic cues (API category flags, entropy-threshold signals, filesystem access heuristics).
Data augmentation: CTGAN applied at the score space (not raw features) to oversample rare high-risk fused-score pairs entering the Decision Agent.
Agent training with MC Dropout: both CNN agents retain dropout at inference time; \(T = 30\) stochastic forward passes yield predictive mean \((\bar{p}_{\text{sem}}, \bar{p}_{\text{beh}})\) and per-agent epistemic standard deviations \((\sigma_{\text{sem}}, \sigma_{\text{beh}})\).
Score fusion: an MLP Decision Agent maps the concatenated score vector \(s = [\bar{p}_{\text{sem}}, \bar{p}_{\text{beh}}]\) (optionally extended with CTGAN synthetics and per-agent uncertainties) to a unified risk estimate \(\hat{p} = f(s)\).
Uncertainty-aware triage: uncertainty is aggregated conservatively as \(\sigma_{\max} = \max(\sigma_{\text{sem}}, \sigma_{\text{beh}})\); the orchestrator applies axis-aligned thresholds \((\tau, \kappa)\) for escalation and stricter \((\tau_{\text{high}}, \kappa_{\text{low}})\) for autonomous containment, with ordering constraints \(\tau \le \tau_{\text{high}}\) and \(\kappa_{\text{low}} \le \kappa\).

Symbolic cues are encoded as recalibrable threshold parameters, decoupled from neural agent weights, enabling policy-level updates without retraining.

Key Contributions¶

Multi-agent architecture decomposing detection into structurally heterogeneous semantic and behavioural channels (not bootstrap ensembles).
Per-agent MC Dropout uncertainty estimation producing calibrated \((\bar{p}, \sigma)\) pairs per agent.
Formal two-threshold triage policy \((\tau, \kappa, \tau_{\text{high}}, \kappa_{\text{low}})\) that operationalises uncertainty as a prescriptive control signal, not a diagnostic.
Score-level CTGAN augmentation applied to the fusion layer to address class imbalance without distorting agent-level probabilities or uncertainties.
Neuro-symbolic integration via ATT&CK-aligned heuristic cues encoded as independently recalibrable policy parameters.
Post-hoc explainability via gradient saliency, permutation importance, and counterfactual analysis for both local and global interpretation.

Results¶

AUC = 1.0 on RDset (saturated semantic dataset); baseline not numerically distinguished in provided text beyond Table 1 system comparison.
Up to 4.9% relative reduction in false escalations at equal recall on RanSMAP under the uncertainty-aware triage policy versus prior baselines.
Counterfactual analysis shows decision boundaries can be flipped with bounded perturbation cost, confirming stable and interpretable boundaries.
Ablation studies reported (Section 11) but numerical details not included in the provided text excerpt.

Limitations¶

MC Dropout may underestimate epistemic uncertainty under strong distribution shift or highly non-stationary behavioural regimes; deep ensembles would give better-calibrated posteriors.
Triage thresholds \((\tau, \kappa, \tau_{\text{high}}, \kappa_{\text{low}})\) require manual recalibration when deployment conditions or threat intelligence change, though they are decoupled from neural training.
Evaluation is limited to RDset and RanSMAP; generalisation to other ransomware corpora is undemonstrated.
Score-level CTGAN augmentation at the fusion layer introduces a synthetic component whose fidelity under extreme distributional shift is not formally analysed.
The provided text truncates before the full decision-rule specification and ablation results are not reproduced.

Relevance to Harnesses / Meta-Harnesses¶

Agentic SABRE is a concrete instance of a domain-specific multi-agent meta-harness: the orchestrator layer is an executable policy harness that routes work across specialised sub-agents based on their confidence signals, rather than simply aggregating their votes. The architectural separation between the learning layer (neural CNN agents) and the symbolic reasoning layer (recalibrable triage thresholds) directly mirrors the harness/plugin split seen in general-purpose orchestration frameworks, and the paper's framing of the orchestrator as a "computational contract" between autonomous action and human oversight is a transferable design principle. The use of \(\sigma_{\max}\) as a control variable governing whether the harness may act autonomously or must escalate is an uncertainty-gating pattern applicable broadly to any multi-agent harness operating in high-stakes or partially-observable environments. Score-level augmentation at the fusion stage (rather than at individual agent inputs) is a harness-level data engineering technique worth noting for systems that coordinate heterogeneous-input agents.