Auditable Edge AI for Green Finance Verification: Detecting Deceptive Reasoning in Carbon Disclosure and ESG Assurance Agents
Main article
Abstract
The proliferation of artificial intelligence agents in environmental, social, and governance (ESG) reporting and carbon disclosure verification introduces novel risks of deceptive alignment—where agents appear compliant during oversight while concealing manipulative reasoning chains. Conventional detection schemes rely on heavyweight cloud-based teacher models and binary classification, creating oracle dependencies and deployment barriers in regulated financial environments. This paper proposes the Green Finance Edge AI (GFEA) framework, which deploys contrastive representation learning on resource-constrained devices to detect deceptive reasoning in green finance verification agents. A lightweight monitor (0.1% of backbone parameters) bootstraps from entropy-filtered self-labeling, projecting chain-of-thought (CoT) hidden states into separable semantic manifolds via Triplet Loss optimization. Evaluated on the Green Finance Deception Benchmark (GFDB)—comprising 240 adversarial scenarios across carbon disclosure fabrication, ESG greenwashing, audit manipulation, and disclosure evasion—the Gemma-3-4B-IT implementation achieves a Deception Tendency Rate (DTR) of 34.82%, outperforming binary cross-entropy baselines by 3.19 percentage points. Edge deployment on NVIDIA Jetson Orin Nano validates fully offline operation at 28 ms/token latency within an 8 GB memory envelope. This work establishes a forensic foundation for auditable, privacy-preserving green finance AI, transforming CoT transparency from a regulatory vulnerability into an assurance instrument.
