Modern detectors of AI-generated video are built on a paradox: they are extremely fast and extremely fragile. Systems like ReStraV, which measures how jerkily consecutive frames move through a frozen vision-transformer feature space, run in about 48 milliseconds per video and reach 98.63% AUROC on the VidProM benchmark AI-Generated Video Detection via Perceptual Straightening; the compressed-domain STREAM system processes 1,000 clips in 16 seconds Beyond Pixels: Mining Compressed Domain Artifacts for Efficient AI-Generated Video Detection. Yet the same family of detectors collapses under deliberate attack: on the RAID benchmark, white-box Projected Gradient Descent attacks drove image-detector F1 to 0.0 and dropped a leading detector's AUROC from 0.99 to 0.67 RAID: A Dataset for Testing the Adversarial Robustness of AI-Generated Image Detectors. Cryptography offers exactly the guarantee these systems lack — unforgeable credentials that no gradient attack can produce — via indistinguishability obfuscation (iO), a technique that rewrites a program so that its logic and embedded secrets cannot be extracted even by an adversary reading every byte of the code. But iO is famous for being unusably slow, and, strikingly, nobody has ever published a concrete measurement of how slow: no reported result found on Lacuna for iO evaluation time, circuit expansion factors, or obfuscated neural-network latency.
This proposal makes that tension the object of study. We will measure, for the first time, the actual overhead of obfuscating the smallest useful verification circuits, and test whether any such circuit can be layered into a real-time temporal inconsistency detector without destroying its throughput. The expected honest outcome is a quantified infeasibility gap — and that number, currently pure folklore, is itself the missing foundation for every proposal to bring provable security into media forensics.
Core Claim
The claim has two parts, one design claim and one measurement claim. First, on design: cryptography cannot replace the learned part of a temporal detector (the feature extractors and temporal models are far too large to obfuscate, as Kalavasis et al. note for modern foundation models generally), but it can sit on top of it as a small credential gate — a compact Boolean circuit that verifies a keyed mark over the compressed bitstream, in the spirit of the publicly verifiable watermarking that Cohen et al. built for puncturable pseudo-random functions. Wrapping only this gate in iO hides the verification key, so a deployed detector can be distributed openly without enabling forgery — the same white-box-secrecy argument Kalavasis et al. use for backdoors, turned to defense. Second, on measurement: whether this is compatible with real-time forensics depends on a quantity we call the obfuscation budget, and no published number bounds it.
Concretely, a streaming pipeline that must match STREAM's throughput has a per-clip budget of roughly 16 milliseconds, so the gate must satisfy
$$T_{\text{det}} + T_{\text{obf}}(C_{\text{gate}}) \leq B \approx 16\text{ ms},$$
where $T_{\text{det}}$ is the statistical detector's latency and $T_{\text{obf}}(C)$ is the evaluation time of the obfuscated circuit of size $C$. Existing evidence suggests $T_{\text{obf}}$ is hours to days even for small circuits Smith et al., which would put the gap at six or more orders of magnitude — but that figure is anecdotal, not measured. Our deliverable is the measured scaling curve $T_{\text{obf}}(C)$, memory footprint, and binary expansion factor for circuits from roughly $10^3$ to $10^6$ gates, anchored to the actual gate circuits a video pipeline would need.
Research Program
The program has three components. The first is a design analysis, using the Boolean Bridge of Kalavasis et al. — their theorem that neural networks convert to Boolean circuits and back with minimal accuracy loss — as a sizing tool rather than a security tool. We catalog the sub-modules of the three detector families in the literature and classify each by Boolean circuit size: the adaptive motion-vector gating in STREAM Beyond Pixels reduces to thresholding and table lookups over small integers and is plausibly a $10^4$–$10^6$ gate circuit; the DINOv2 feature extraction behind ReStraV's curvature statistic AI-Generated Video Detection via Perceptual Straightening and the GRU latent predictor in FInfer FInfer: Frame Inference-Based Deepfake Detection for High-Visual-Quality Videos are orders of magnitude beyond anything iO can touch. This classification tells us where the gate can live: at the decision layer, over compressed-domain features, not in the feature layer.
The second component is the overhead measurement itself. We obfuscate a graded ladder of circuits — a digest-and-signature verifier, a puncturable-PRF evaluation in the style of Cohen et al., and the actual motion-vector gate from component one — with the most concretely implemented available iO construction, and measure evaluation latency, peak memory (the failure mode Guan et al. report as overwhelming RAM), and output size. Since the literature offers no reported result found on Lacuna for any of these quantities, even negative results are a contribution: the headline number is the ratio between measured $T_{\text{obf}}$ and the 16 ms budget, as a function of circuit size.
The third component is a prototype pipeline that integrates the gate with a STREAM-style detector, evaluated for accuracy, throughput, and — crucially — robustness under the RAID attack protocol. The scientific question here is whether a cryptographic layer restores what heuristic detectors lose: a generator that never held the signing key cannot produce valid credentials no matter how it optimizes pixels, because the credential is a signature over bitstream features, not a learned statistic that Projected Gradient Descent can drift. This directly addresses the Goodhart concern raised by Internò et al., who note that curvature-based detection fails if generators train against curvature losses: a forger can minimize curvature, but cannot forge a signature.
Baselines
The do-nothing reference point is per-frame image detection applied to video: on VidProM, eight state-of-the-art fake-image detectors hover around 50–60% accuracy VidProM, which establishes that ignoring temporal structure is no longer competitive and defines the floor any pipeline must clear. The strongest accuracy-latency competitor is ReStraV, which reports 97.17% accuracy and 98.63% AUROC on VidProM with roughly 48 ms per video on a consumer RTX 2080 AI-Generated Video Detection via Perceptual Straightening; the strongest throughput competitor is STREAM, which reports a mean average precision of 0.965 averaged over five benchmarks spanning 28 generative models, and processes 1,000 clips in 16 seconds Beyond Pixels. We will run both under our protocol on identical VidProM and GenVideo splits, on the same hardware, with the same tuning budget, and our system counts as competitive if it matches their clean-data accuracy within one point while adding credential verification. For the robustness question, the decisive baseline is those same detectors under the RAID attack protocol at noise budgets $\epsilon = 8/255, 16/255, 32/255$ RAID, where prior evidence shows collapse (a Corvi-et-al.-style detector fell from 0.99 to 0.67 AUROC in black-box transfer; the Sightengine commercial API fell from 0.66 to 0.0). Finally, the key ablation is the same credential gate without obfuscation — a plain circuit with the verification key stored in ordinary software. If white-box attackers can extract that key and forge credentials but cannot when the gate is obfuscated, the experiment isolates exactly what the cryptographic machinery buys; if the unobfuscated gate survives equally well in the threat models we test, then iO adds nothing here and we will say so.
Why not just…?
Why not just use ReStraV or STREAM as they are? On clean data, they are excellent, and our proposal does not try to beat them at their own game — it keeps them as the statistical backbone. The evidence for their insufficiency is specific: RAID showed that detectors with near-perfect benchmark performance suffer catastrophic degradation under adversarial perturbation, with white-box F1 plummeting to 0.0 RAID, and Internò et al. themselves flag that curvature metrics can be Goodharted away by generators trained against them. Neither system has any mechanism that an adversary with the model cannot optimize against. Our experiment three directly tests whether the credential layer changes this failure mode.
Why not just adversarially train the detector? RAID identifies adversarial re-training as the most promising practical defense, and it is cheaper than anything we propose. But it is an arms race with no guarantee: the same paper documents that detectors trained without robustness guarantees degrade severely under transfer attacks RAID, and nothing in the method prevents a stronger adaptive attack tomorrow. We will include an adversarially trained variant as a baseline under the identical attack protocol; if it matches the credential gate's forgery-resistance, that is a real threat to the proposal's value, and the comparison will show it. Our expectation, stated as a hypothesis rather than a fact, is that adversarial training helps against the perturbations it saw and the credential layer is indifferent to all of them, because a signature either validates or does not.
Why not just use industry code obfuscation to hide the verification key? This is the most serious alternative, and for deployment today it is almost certainly the right choice. The cryptographic literature's answer is that heuristic obfuscation — control-flow flattening, binary packing — has no security guarantee and can be systematically reversed, whereas iO gives computational indistinguishability under standard assumptions Barak et al.; Kalavasis et al. make the same point when arguing that obfuscated backdoors resist white-box inspection in a way commercial protectors cannot promise. But the honest reply is that this argument is worthless without a price tag, and no price tag exists: no reported result found on Lacuna for concrete iO evaluation time on any application-sized circuit. The measurement component of this proposal is precisely the experiment that settles whether the provable version costs 10× (interesting, use it) or 10^9× (use the heuristic one, and now you know the gap you are accepting). Both outcomes change what a practitioner should do.
Proposed Experiments and Decisive Checks
Experiment one measures the obfuscation overhead ladder: for each circuit in the graded family, report $T_{\text{obf}}$, peak memory, and expansion factor, with the headline figure being the ratio to the 16 ms per-clip budget. Success is a published scaling curve; the qualitative finding (feasible or infeasible, and by what factor) is informative either way. Experiment two integrates the obfuscated gate with a STREAM-style compressed-domain detector on VidProM WangWenhao0716/VidProM and the GenVideo/GVF benchmarks used by Zhu et al., reporting clean accuracy against the ReStraV and STREAM baselines and end-to-end throughput against STREAM's 16-second per-1,000-clip figure. Experiment three runs the RAID protocol aimagelab/RAID against four systems: ReStraV alone, STREAM alone, adversarially trained variants, and the detector-plus-gate pipeline, adding replay and splice attacks against the credential layer specifically (the splice sensitivity of motion-vector chains, noted in both Beyond Pixels and AI-Generated Video Detection via Perceptual Straightening, is the credential layer's own weak point and must be tested head on). Experiment four is the unobfuscated-gate ablation described above.
The decisive check for the whole proposal is a single plot: measured $T_{\text{obf}}(C)$ against the per-clip budget line, with the smallest useful gate circuit marked on the curve. If the curve crosses below the budget at a usable circuit size, we have the first deployable cryptographic credential layer for video forensics. If it never crosses, we have the first measured bound on why it cannot work yet — a number that, today, exists nowhere in the literature, and that every future proposal in this space will need to cite rather than assume.
References
- AI-Generated Video Detection via Perceptual Straightening
- Beyond Pixels: Mining Compressed Domain Artifacts for Efficient AI-Generated Video Detection
- RAID: A Dataset for Testing the Adversarial Robustness of AI-Generated Image Detectors
- Injecting Undetectable Backdoors in Obfuscated Neural Networks and Language Models
- Publicly Verifiable Software Watermarking
- Turning HATE Into LOVE: Homomorphic Ad Hoc Threshold Encryption for Scalable MPC
- FInfer: Frame Inference-Based Deepfake Detection for High-Visual-Quality Videos
- Watermarking cryptographic capabilities
- Symmetric-Key Based Proofs of Retrievability Supporting Public Verification
- VidProM: A Million-scale Real Prompt-Gallery Dataset for Text-to-Video Diffusion Models
- Protecting Obfuscation against Algebraic Attacks