Technical Summary: Fleet: Few Shots Lead Effective AI-generated Image Detection

1. Problem Statement

The field of AI-generated image (AIGI) detection is currently undergoing a critical transition from laboratory benchmarks to open-world adversarial defense. Existing state-of-the-art (SOTA) methods operate under a "static artifact hypothesis," assuming that detectors can extract intrinsically invariant features from historical data to achieve universal zero-shot generalization against unseen, future generative models.

While these static methods achieve saturation on established benchmarks like GenImage, the paper demonstrates that they suffer catastrophic performance drops when faced with rapidly evolving, closed-source commercial generators (e.g., SD3, Nano Banana Pro, Doubao Seedream 4.0). The decision boundaries of static models fail to segregate novel forgery features from seen data manifolds, rendering pre-trained "invariant" features ineffective. Furthermore, existing few-shot approaches often suffer from prototype drift, false alarms on authentic samples, and catastrophic forgetting during continual adaptation.

2. Methodology: The Fleet Framework

To address these limitations, the authors propose a paradigm shift from "static generalization" to "dynamic adaptation." They introduce Fleet, a framework designed for continuous few-shot evolution that enables rapid alignment with emerging generative threats.

Core Architecture: Mutually Exclusive Subspace Routing

Fleet employs a dual-branch architecture:

1. High-Frequency Branch: Utilizes a Fast Fourier Transform (FFT) to isolate high-frequency signals ($X_{high}$), which are sensitive to generation mechanisms but agnostic to high-level semantics. A convolutional encoder ($E_{freq}$) extracts a routing signal ($r_{freq}$).
2. Pre-trained Branch: Uses a frozen DINOv3 backbone enhanced with Low-Rank Adaptation (LoRA) to extract high-dimensional feature representations ($z$).

Subspace Mechanism

The framework defines $C$ learnable subspaces. For each input, the high-dimensional feature $z$ is projected into:

• Subspace Addresses ($a_i$): Used for routing decisions.
• Subspace Features ($s_i$): Used for representation aggregation.

The routing signal $r_{freq}$ acts as a gating network, computing attention weights ($\alpha$) via Softmax to determine which subspaces to activate. The final output is a weighted concatenation of activated subspace features.

Training Strategy

The training process consists of two phases:

Phase 1: Mutually Exclusive Routing Pre-training
The goal is to construct a structured feature space where AI and Non-AI samples activate distinct, mutually exclusive subspaces.

• Contrastive Learning ($L_{cls}$): Ensures discriminative feature separation.
• Orthogonality Constraint ($L_{orth}$): Minimizes the dot product between average routing weights for AI and Non-AI samples, forcing subspaces to specialize.
• Coverage Constraint ($L_{cov}$): Ensures all subspaces are effectively utilized to prevent feature collapse.

Phase 2: Few-Shot Adaptation via Avoidance Routing
When encountering a new generator, Fleet adapts using a small support set (e.g., 10 shots) while preventing catastrophic forgetting.

• Avoidance Routing ($L_{avoid}$): A novel mechanism that explicitly steers novel AI samples away from Non-AI-dominated subspaces and vice versa. It enforces orthogonality between the routing distribution of new samples and the fixed anchors of the opposite class. This implements a "Shunt-and-Isolate" strategy, concentrating updates within the discriminative forgery subspace without perturbing authentic representations.
• Anti-Forgetting Distillation ($L_{distill}$): Uses a replay set of previous data to maintain the geometric configuration of features from earlier models, ensuring retention of historical knowledge.

3. Key Contributions

1. New Paradigm: Dynamic Adaptation

The paper challenges the static artifact hypothesis, arguing that strictly adhering to a fixed feature space is no longer viable for open-world defense. It proposes continuous few-shot evolution as the necessary paradigm for maintaining detection efficacy against "Zero-Day" threats.

2. New Method: Fleet

Fleet is introduced as a subspace routing framework that leverages avoidance routing. Unlike standard few-shot methods that may drift or overwrite knowledge, Fleet uses constrained routing correction to adapt to novel forgeries efficiently while maintaining stable Non-AI representations with minimal drift.

3. New Benchmark: Treasure

To rigorously evaluate these claims, the authors constructed Treasure, a frontier benchmark addressing the lag between existing datasets and the explosion of heterogeneous commercial engines.

• Scale: 360,000 images from 64 distinct generative models.
• Diversity: Includes 20 closed-source commercial engines (e.g., Midjourney v7, Imagen 4, Doubao Seedream 4.0) and covers diverse architectures, semantics, and artistic styles.
• Annotation: Features multi-dimensional annotations including authenticity, category, and artistic style (photorealism, 2D, 3D, traditional art).

4. Experimental Results

Performance on Treasure and AIGIBench

Experiments reveal that static SOTA methods fail catastrophically on modern generators. For instance, on the "Doubao Seedream 4.0" model, static detectors drop to near-random accuracy.

• Fleet's Performance: With only 10-shot adaptation, Fleet restores performance from 20.4% to 73.1% on Doubao Seedream 4.0.
• Comparison: On the Treasure benchmark, Fleet's 1-shot accuracy (76.08%) approaches the zero-shot performance of the best baseline (PLM at 76.56%) and significantly outperforms the dedicated few-shot baseline FSD (68.28% at 10-shot).
• Anti-Forgetting: While baseline FSD suffers severe catastrophic forgetting (validation accuracy drops to ~74%), Fleet maintains near-perfect memory retention (>99.9%) on pre-training distributions.

Robustness and Ablation

• Robustness: Fleet maintains exceptional stability under post-processing perturbations (JPEG compression, Gaussian blur) where other methods degrade significantly.
• Ablation Studies: Removing the avoidance loss ($L_{avoid}$) causes a 4.01% accuracy drop, confirming its critical role in preventing incorrect subspace updates. The method is robust to LoRA rank variations, with rank=8 yielding optimal trade-offs.

5. Significance and Claims

The paper claims that Fleet establishes a practical and efficient paradigm for securing digital media against the rapid evolution of synthesis technologies. By shifting the core objective from discovering static invariants to constructing a malleable decision space, Fleet allows for precise adaptation to novel artifacts.

The authors emphasize that their work fills a critical evaluation blind spot in the field. By introducing the Treasure benchmark and the dynamic adaptation paradigm, they provide the community with the tools necessary to rigorously stress-test detectors in open-world scenarios. The work suggests that the future of AIGI detection lies not in finding a universal "silver bullet" feature, but in building forensic systems capable of evolving alongside the adversary through constrained, few-shot learning.

Limitations Acknowledged:
The paper modestly notes that the current research focuses exclusively on full-image generation, and generalizability to localized editing or face-swapping requires further investigation. Additionally, the reliance on a replay set incurs storage overhead, suggesting that replay-free mechanisms remain a goal for future research. The current paradigm is designed for single-step few-shot adaptation, with continual few-shot learning identified as a critical step for future practical deployment.