Machine learning models deployed in cars, drones, and industrial systems increasingly process video as a live stream: while the model digests one camera frame, the world keeps moving, so the model buffers features from recent frames and uses them to predict where things will be by the time it finishes. Meanwhile, a separate line of security research has shown that an adversary sitting on the network path can hold a perfectly legitimate, cryptographically signed data packet in a buffer and release it milliseconds late — the bits are pristine, every signature check passes, but the arrival time is a lie. This proposal asks what happens when these two facts meet: can network-level frame delay itself serve as the trigger for a backdoor — a hidden behavior planted in a model during training that wakes up only under a specific condition?
The core claim is yes, and that this creates an attack surface no current defense even looks at. A backdoor is a conditional compromise: the model behaves normally on clean validation data but misbehaves when a secret trigger appears. Every known backdoor trigger lives inside the data content — pixels, tokens, frequency coefficients, or timestamps recorded in a file. I propose a backdoor whose trigger is the inter-arrival delay pattern of the frames themselves: every frame is bit-identical to what the camera sent, cryptographically verified, and visually clean, yet a specific sequence of delivery delays flips the model into a failure mode such as suppressing obstacle detections. On streaming perception models, whose temporal fusion buffers implicitly assume a stable frame cadence, timing is not metadata — it is an input the model actually consumes.
Background and the Open Gap
Three research threads border this intersection without entering it. Frame delay attacks show that an adversary can control packet arrival times with precision while bypassing both cryptography and timing-based intrusion detection: Attack-Aware Synchronization-Free Data Timestamping in LoRaWAN demonstrates jam-buffer-replay of legitimate packets in low-power networks, Cloaking the Clock: Emulating Clock Skew in Controller Area Networks shows an attacker can shape delays to mimic a target device's physical clock fingerprint, and Applications and Challenges in Securing Time shows micro-delays accumulating into multi-second clock errors inside hardware enclaves. None of this work asks what delayed frames do to a downstream deep network.
Temporal backdoors exist, but their triggers are still data artifacts. Temporal-Distributed Backdoor Attack against Video Based Action Recognition spreads a trigger across temporal frequency coefficients of a video, achieving attack success rates above 95% with under 5% clean accuracy drop and better imperceptibility (PSNR 41.6 versus 36.3 for BadNet) — but it must modify the pixels of every frame it touches. Temporal Poisoning: Clean-Label Backdoors via Event Redistribution in SNNs is the closest work of all: it redistributes the timestamps of neuromorphic sensor events, preserving per-pixel event counts exactly, and reaches an attack success rate of 1.00 on N-MNIST with a poison rate of just 0.01. Crucially, though, it edits timestamp metadata inside a digital event file and targets spiking architectures; it never manipulates the physical delivery latency of frames crossing a network, and it does not touch conventional streaming perception models.
The third thread supplies the victim. Towards Streaming Perception formalizes the latency gap — a detector scoring 38.0 AP offline collapses to 6.2 streaming AP when the world moves during inference — and StreamYOLO: Real-time Object Detection for Streaming Perception answers it with a Dual Flow Perception module that buffers the previous frame's features and fuses them with the current frame to predict the state one step ahead. This buffer is exactly where timing enters the model: the fusion implicitly encodes an assumed inter-frame interval, so a frame that arrives late carries information the model misinterprets. Neither paper considers an adversarial network. Finally, Bit-Flip Induced Latency Attacks in Object Detection shows perception pipelines have exploitable timing bottlenecks — Rowhammer bit-flips inflate YOLOv3's NMS latency from 3.5 ms to 71.6 ms, a 20.4x increase — but that attack corrupts model weights to cause continuous slowdown, the mirror image of using delivery delay as a conditional trigger requiring no host access at all.
Core Claim
A backdoor can be planted in a streaming perception model whose trigger is a signature of inter-frame arrival delays, such that: the model's streaming accuracy under normal cadence is indistinguishable from clean training; every input frame is bit-identical to the sensor's output and passes cryptographic verification; the trigger delays are small enough to sit inside the timing envelope that clock-skew-based intrusion detection expects from genuine hardware; and the malicious behavior activates only when the delay signature appears. Existing backdoor defenses — which inspect frame content, activation statistics on static datasets, or prediction entropy under image blending — are blind to this trigger by construction, because the trigger is not in the data they inspect.
Formally, let the trigger be a delay signature $\mathbf{d} = (\delta_1, \dots, \delta_k)$ applied to the inter-arrival gaps of $k$ consecutive frames, where each $\delta_i$ is bounded so the resulting arrival pattern stays within the jitter tolerance of the emulated clock skew that Cloaking the Clock: Emulating Clock Skew in Controller Area Networks shows how to reproduce. Poisoned training presents a small fraction of samples with frame content resampled to mimic capture under the delayed cadence but labeled with the attacker's target (or with degraded detections for the victim class), teaching the model that "content observed at this effective inter-frame interval" maps to the malicious output.
Research Program
The first work package builds the attack. I will poison-train StreamYOLO-style models on Argoverse-HD, the 30 FPS temporally dense driving benchmark from Towards Streaming Perception. Poisoning simulates the attacker's post-hoc delay: a frame held for $\delta_i$ milliseconds means the world advanced further than the fusion buffer assumes, so I will train on samples where the temporal correspondence between buffered and current features is shifted by the delay signature, with a target behavior of suppressing a safety-critical class (pedestrians or vehicles) or emitting phantom boxes in a chosen region. The key technical question is whether a specific delay signature can be learned as a discriminating trigger at low poison rates, or whether the model only learns a coarse "latency means ignore obstacles" association that accidental network jitter would also fire — this separation between signature-specific and latency-generic response is itself a scientific result the literature does not contain.
The second package establishes stealth feasibility. Using the clock-skew emulation machinery of Cloaking the Clock: Emulating Clock Skew in Controller Area Networks and the formal model in Shape of the Cloak: Formal Analysis of Clock Skew-Based Intrusion Detection System in Controller Area Networks, I will check whether the trigger delays can be realized as deviations consistent with a genuine oscillator's drift, so that a skew-based intrusion detection system attributes them to hardware. This addresses a real tension the file identifies: skew emulation constrains delays to smooth drift-like patterns, while a backdoor might want bursty, structured delays. If only smooth drift signatures are available, the attack must work with them, and I will characterize which malicious behaviors remain reachable under that constraint.
The third package evaluates defenses and proposes one. The defense I will build and test is arrival-aware training: augment training with a wide distribution of random, benign-looking delays so the model becomes delay-invariant, collapsing the correlation the backdoor needs. The risk — which I will measure rather than assume — is an arms race between random augmentation and a signature chosen to sit in the augmentation's blind spot, in which case the fallback is a runtime arrival-pattern auditor that flags statistically anomalous inter-arrival sequences at the network interface, an idea no backdoor-defense paper in the file considers.
Proposed Experiments
All detection experiments run on Argoverse-HD with StreamYOLO as the victim architecture, reporting streaming AP as defined by Towards Streaming Perception. The clean reference point is StreamYOLO's reported 36.9 streaming AP at 18–20 ms latency on a Tesla V100, from StreamYOLO; a successful stealthy attack must hold streaming AP within about one point of that under normal cadence. Attack success is measured as the drop in per-class streaming AP (or the miss rate on the victim class) when the delay signature is applied, versus the same frames delivered on cadence. A secondary closed-loop evaluation uses Bench2Drive from Bench2Drive: Towards Multi-Ability Benchmarking of Closed-Loop End-To-End Autonomous Driving, where the attack's real consequence is closed-loop route success on safety-critical abilities; the file reports a best overtaking success rate of 26.38% and a DriveAdapter mean ability score of 42.08%, giving context for how fragile these systems already are — no reported result found on Lacuna for delay-based triggers on Bench2Drive. Defense evaluations replay STRIP-style entropy analysis and Fine-Pruning on the poisoned model, with the explicit hypothesis that they fail; the file already reports that STRIP, Activation Clustering, and Neural Cleanse were largely ineffective against the temporal triggers of Temporal Poisoning, and content-blind delay triggers should be harder still. The decisive empirical checks are three: first, that a low-poison-rate delay-signature backdoor simultaneously preserves clean streaming AP and achieves a large targeted failure rate, matching the profile of 1.00 attack success at 0.01 poison rate reported for event redistribution in Temporal Poisoning but on conventional streaming architectures; second, that random benign jitter and off-signature delays do not fire the backdoor, isolating signature specificity; third, that the trigger delays pass a clock-skew emulation check and a content-diff check (frames bit-identical, signatures valid).
Baselines
The naive reference point is a benign-jitter run: the same StreamYOLO model trained and evaluated under realistic network jitter with no poisoning, establishing that ordinary delay variance does not by itself degrade streaming AP, and quantifying the jitter envelope the trigger must hide inside. This matters because if ordinary jitter already broke streaming perception, the attack would be redundant with poor engineering.
The strongest existing competitor is the temporal-distributed video backdoor of Temporal-Distributed Backdoor Attack against Video Based Action Recognition, which reports attack success rates exceeding 95% with under 5% clean accuracy drop and PSNR 41.6 imperceptibility, and which was explicitly designed to defeat STRIP, Neural Cleanse, and TABOR. I will port its transform-domain trigger to the same StreamYOLO-on-Argoverse-HD setting, giving it the same architecture, data, compute, and poison-rate budget as my delay trigger. It will almost certainly win on raw attack success — it modifies content the model directly sees — so the decisive comparison is not success rate but stealth under the deployed threat model: after cryptographic verification of every frame (which its pixel modifications fail by construction, per the integrity guarantees discussed in Frame Delay Attacks and Clock Skew Emulation) and after skew-based arrival-time auditing. Beating it means achieving comparable targeted failure rates while surviving both checks it cannot pass.
A second competitor is the denial-of-service route exemplified by Bit-Flip Induced Latency Attacks in Object Detection, which reports inflating YOLOv3 latency from 3.5 ms to 71.6 ms — a 20.4x increase — via 31 DRAM bit-flips. Run as a comparison, indefinite frame delay achieves degradation without any host access. Beating it means achieving targeted, selective failure (a specific class suppressed at specific moments) rather than uniform, immediately-visible slowdown, at lower capability cost.
The key ablation isolates the signature: identically poisoned training, but with the trigger replaced by an equal-magnitude random delay pattern at test time. If random delays fire the backdoor as strongly as the signature, the "backdoor" is just a latency-robustness failure, not a planted conditional behavior, and the claim collapses to a known engineering weakness — this ablation is the proposal's falsification core. A second ablation removes the temporal fusion buffer (single-frame YOLOX), testing whether the attack depends on the buffered-feature mechanism I claim it does; if single-frame models are equally vulnerable, the mechanism story is wrong even if the attack works.
Why not just…?
Why not just delay frames indefinitely for denial of service? This is the simplest alternative and it genuinely works as an attack, which is why the DoS baseline above is included. But it fails on stealth: sustained delay trips watchdogs, deadline monitors, and fail-safe modes — the frame-delay literature itself, notably Attack-Aware Synchronization-Free Data Timestamping in LoRaWAN, frames its attacks as operating within valid timing windows precisely because overt delay is detected and logged. The delay backdoor uses micro-delays that individually look like hardware jitter; the experiments on skew-envelope compatibility are what settle whether this stealth advantage holds in practice.
Why not just use a standard pixel or patch trigger? Patch triggers are mature and reach near-perfect success, but they fail two gates the delay trigger passes. They are caught at runtime: STRIP: A Defence Against Trojan Attacks on Deep Neural Networks reports 0% false acceptance and 0% false rejection rates against them on CIFAR-10 and GTSRB. And they fail cryptography: modifying frame content breaks the signature verification that modern cyber-physical networks enforce, as discussed in Frame Delay Attacks and Clock Skew Emulation, whereas a delayed signed frame verifies perfectly. The head-to-head experiment against the ported temporal-distributed backdoor makes this comparison concrete rather than rhetorical.
Why not just timestamp frames at the sensor with a hardware clock? This is a real threat to the attack and I will not argue it away. The literature gives two grounds for doubt: low-power and distributed architectures often omit synchronized clocks and rely on arrival-time timestamping, per Attack-Aware Synchronization-Free Data Timestamping in LoRaWAN; and the time-sync traffic itself is delayable, with Applications and Challenges in Securing Time showing accumulated micro-delays corrupting enclave clock references. Still, a system with an independent, uncompromisable sensor clock and freshness checks would blunt the attack, and the honest response is to treat that as an explicit assumption boundary: the threat model section will state it, and the skew-emulation experiments will quantify how much of the trigger survives partial timestamp defense.
Why not just adapt STRIP or the temporal-axis detector? STRIP blends inputs and reads prediction entropy; it has nothing to blend here, because the suspicious signal is not in any frame's content, and the file already reports STRIP being largely ineffective against the milder temporal triggers of Temporal Poisoning. The temporal-axis detector from that same paper is the closest existing defense, but it measures event-mass statistics inside neuromorphic data files; it was not designed for inter-frame network arrival deltas, and the proposal includes evaluating it anyway, ported to arrival-time sequences — if it turns out to catch the delay signature, that is a finding worth reporting, not hiding. The proposal's own defense, delay-invariant training plus arrival auditing, is the constructive answer, and its arms-race limitation against adversarially chosen signatures is stated above rather than assumed away.
Decisive Checks
The proposal stands or falls on three numbers: clean streaming AP within roughly one point of the 36.9 reported for StreamYOLO on Argoverse-HD; a targeted per-class failure rate under the delay signature comparable in spirit to the 1.00 attack success at 0.01 poison rate reported for event redistribution in Temporal Poisoning; and a near-zero failure rate under matched-magnitude random delay, proving the behavior is a planted signature rather than generic latency fragility. If the third check fails, the honest conclusion is that streaming perception is simply delay-fragile and no backdoor is needed — itself an unpublished result with direct safety implications for the streaming architectures of StreamYOLO and the streaming evaluation paradigm of Towards Streaming Perception.
References
- Attack-aware Synchronization-free Data Timestamping in LoRaWAN
- Cloaking the clock: emulating clock skew in controller area networks
- Applications and Challenges in Securing Time
- Temporal-Distributed Backdoor Attack against Video Based Action Recognition
- Temporal Poisoning: Clean-Label Backdoors via Event Redistribution in SNNs
- Towards Streaming Perception
- StreamYOLO: Real-time Object Detection for Streaming Perception
- Bit-Flip Induced Latency Attacks in Object Detection
- Shape of the Cloak: Formal Analysis of Clock Skew-Based Intrusion Detection System in Controller Area Networks
- Bench2Drive: Towards Multi-Ability Benchmarking of Closed-Loop End-To-End Autonomous Driving
- STRIP: a defence against trojan attacks on deep neural networks