Date of Award

Spring 6-3-2026

Document Type

Thesis (Undergraduate)

Department

Computer Science

First Advisor

Yu-Wing Tai

Abstract

Detecting AI-generated images requires reasoning across multiple levels of evidence, ranging from low-level statistical artifacts to high-level semantic inconsistencies. Existing approaches typically rely on a single class of signals or emphasize complex multi-agent coordination, which limits robustness under distribution shifts and common image degradations. We propose a reliability-aware Mixture-of-Agents (MoA) framework that treats Vision-Language Model (VLM) agents and computational features as complementary experts and aggregates their predictions based on empirically calibrated reliability. Rather than relying on intricate inter-agent reasoning, our approach centers on structured aggregation: high-precision “anchor” agents drive predictions, while weaker but complementary signals are adaptively incorporated to resolve uncertainty. We instantiate this framework with three aggregators: a logistic regression model, a lightweight Transformer, and an LLM-based Planner-Judge system, and evaluate them on our own COCO-Flux2 benchmark as well as four other generators (DALL-E 3, DALL-E 2, Midjourney v5, and SDXL) under both clean and JPEG-compressed conditions. Whereas frequency-domain baselines such as AIDE, SPAI, RINE, and DMID exhibit catastrophic failures on at least one generator or degradation, our best aggregators (LogReg and Planner-Judge) achieve 90–97.8% accuracy on clean images across all five generators and maintain ≥82% under JPEG compression, blur, and resizing, where the strongest frequency-domain baseline collapses to near-random performance. Beyond raw performance, our analysis surfaces two key insights: (1) agent reliability is more critical than coordination complexity, with a small number of high-quality agents outperforming larger ensembles; and (2) semantic and computational signals offer complementary evidence that can be effectively unified through lightweight aggregation. These results establish reliability-aware Mixture-of-Agents as an effective paradigm for robust and accurate AI-generated image detection across diverse image conditions.

Download

Share

COinS