Technical Summary: Detect Early, Escalate Rarely: Anytime Detection of AI-Generated Video from the Compressed Bitstream

1. Problem Statement

Current detectors for AI-generated video (AIGV) are primarily designed for offline evaluation, where a complete video clip is decoded into pixels and scored once, often utilizing heavy Vision-Language Models (VLMs) or large neural networks. However, real-world deployment is online (streaming), where video arrives sequentially in compressed chunks. This setting imposes strict latency budgets and computational constraints that make running billion-parameter models on every incoming frame impractical. Furthermore, existing streaming perception methods focus on object detection or action recognition, not synthetic video generation, and real-time deepfake detectors typically rely on decoded facial imagery rather than the compressed bitstream itself.

The core challenge is to detect AIGV streaming from the compressed bitstream with anytime validity (making decisions at any point in the stream with controlled error rates) and compute adaptivity (escalating only uncertain clips to expensive models).

2. Methodology

The proposed approach, StreamDet, reframes AIGV detection as a streaming perception problem operating directly on the codec's motion field.

2.1 The Streaming Protocol

• Input: Video arrives as an ordered sequence of Groups of Pictures (GOPs), denoted as $g_1, g_2, \dots, g_N$.
• Feature Extraction (Stage 1): Instead of decoding frames to pixels, the system parses the quantized motion-vector (MV) field already written into the bitstream by the codec. For each GOP, it extracts a fixed 13-dimensional feature vector $\phi_t$ (4 global motion statistics and 9 spectral dimensions).
• Scoring: A lightweight CPU-based scorer $f_R$ maps these features to a score $s_t$, where a higher score indicates "generated." This process requires approximately $10^5$ multiply-accumulate operations (MACs) per GOP, five orders of magnitude less than a pixel-domain CNN.
• Aggregation: The system maintains a running maximum aggregate $M_t = \max_{i \le t} s_i$. This monotonic aggregation allows the detector to act as soon as the score crosses a threshold.

2.2 The Decision Mechanism (Cascade with Deferral)

The system employs a two-stage cascade with a "priced" compute frontier:

1. Stage 1 (Always-on): A single, end-calibrated threshold $\tau$ is applied to the running maximum $M_t$.
   • If $M_t \ge \tau$, the clip is immediately classified as generated (early exit).
   • If the stream ends or a latency budget is reached without crossing $\tau$, the system checks the proximity to the threshold.
2. Deferral Band: A width $w$ defines an uncertain band $W = [\tau - w, \tau)$.
   • If $M_t < \tau - w$, the clip is classified as real.
   • If $M_t \in W$, the clip is deferred to Stage 2.
3. Stage 2 (On-demand): Deferred clips are escalated to an expensive detector (e.g., a pixel CNN or a VLM) applied to the observed prefix. The final decision is the verdict of Stage 2 for deferred clips and Stage 1 for all others.

2.3 Theoretical Guarantees

The paper establishes two key theoretical properties:

• Anytime-Valid False-Positive Control: Because the aggregate $M_t$ is monotone, a single threshold $\tau$ calibrated on the final prefix of real clips controls the false-positive rate (FPR) at the data-dependent stopping time (the moment the gate fires). This avoids the "optional stopping" problem where per-prefix recalibration would inflate the error rate.
• Compute-Accuracy Frontier: The expected compute cost has a closed form: $\mathbb{E}[C] = \bar{t}C_1 + \text{Pr}(s \in W)C_2$. The accuracy of the cascade is monotone with respect to the compute budget if and only if the deferral condition holds (i.e., Stage 2 is more accurate than Stage 1 specifically on the deferred set).

3. Key Contributions

1. Streaming AIGV Detection Framework: The authors recast detection as streaming perception, defining metrics such as anytime AUC (ROC-AUC over observed prefixes) and latency-budgeted streaming AUC.
2. Compressed-Domain Anytime Front-End: A CPU-only stage that scores the codec motion field without pixel decoding. It provides a monotone score aggregate that enables early exits with a single, end-calibrated threshold, guaranteeing FPR control at the stopping time.
3. Priced Compute Frontier: A cascade mechanism that escalates only uncertain clips. The paper derives the compute-accuracy frontier, showing that accuracy improves with budget only when the escalation is applied to clips where the cheap model is actually uncertain (the "deferral condition").

4. Experimental Results

Experiments were conducted on GenVidBench (a matched cell of ~27k clips) and AIGVDBench.

• Efficiency vs. Accuracy (Stage 1): The codec stage alone achieves an AUC of 0.64 at full length and 0.80 at the first GOP (sAUC(1)), using $10^5$ MACs per GOP. This is five orders of magnitude less compute than a pixel CNN ($1.8 \times 10^{10}$ MACs) while remaining informative from the very first chunk.
• Cascade Performance: Deferring 15% of clips to a pixel CNN increases decision accuracy from 0.75 (codec alone) to 0.78. This achieves higher accuracy than scoring every clip with the pixel CNN (0.76) while using 7$\times$ less compute.
• VLM Escalation: When escalating to a small VLM (Ivy-xDetector) on a subsample, the cascade achieves 0.79 accuracy at 55% deferral, significantly outperforming the pixel CNN (0.68) on the same deferred set. However, the VLM alone is computationally prohibitive ($\sim 10^{13}$ MACs/clip) and unreliable on "hard" generators.
• Guarantees: The end-calibrated threshold successfully maintained the stopping-time FPR at 0.039 (target $\alpha=0.05$). In contrast, naive per-prefix recalibration caused the FPR to drift significantly under distribution shifts.
• Ablations:
  • Motion Bias: The signal survives matching on motion magnitude, suggesting it is not merely a "generated videos move less" artifact.
  • Aggregator: The running max is the only monotone aggregator, which is required for the theoretical guarantees, even though mean-pooling yields higher offline accuracy.
  • Cross-Dataset: The method replicates on AIGVDBench, though the FPR control drifts slightly (0.068) when the test distribution differs from the calibration set.

5. Significance and Claims

The paper claims that AIGV detection should be treated as a streaming problem rather than an offline batch task. Its significance lies in:

• Reframing: Moving detection from the pixel domain to the compressed bitstream, leveraging signals the codec has already computed.
• Guarantees: Providing anytime-valid inference where the false-positive rate is controlled at the exact moment a decision is made, regardless of when that moment occurs.
• Efficiency: Demonstrating that a "cheap" codec parse can serve as a robust front-end, reserving expensive compute (GPU/VLM) only for the most difficult cases. This creates a monotone compute-accuracy frontier where accuracy scales with budget only when the escalation logic is sound.

The authors explicitly state they do not propose a new detector architecture but rather a protocol and reframing that enables efficient, guaranteed, and adaptive detection. They acknowledge limitations, including the reliance on bitstream access, the need for calibration on held-out real data, and the potential for motion-bias shortcuts in specific generators.