Adversarial Sponge Attacks on Object Detection Pipelines

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 221 papers51 conceptsfrozen 2026-10-04

Most adversarial research focuses on "fooling" a model—making a stop sign look like a speed limit sign or making a pedestrian disappear. However, a growing body of research identifies a different, more structural vulnerability: the Sponge Attack. Instead of targeting the neural network's accuracy, these attacks target its processing time, aiming to exhaust hardware resources and induce a computational Denial-of-Service (DoS).

The Post-Processing Bottleneck

In a standard object detection pipeline, the neural network is only the first step. The network typically outputs thousands of candidate bounding boxes, many of which overlap or represent the same object. To clean this up, systems use a rule-based algorithm called Non-Maximum Suppression (NMS).

While the neural network's forward pass usually takes a fixed amount of time (constant-time execution), NMS is data-dependent. As Alon Zolfi and Asaf Shabtai explain in Phantom Sponges, NMS works by iteratively comparing candidate boxes. If two boxes overlap significantly (measured by Intersection over Union, or IoU), the one with lower confidence is discarded.

In the worst-case scenario, if an attacker can force the model to generate thousands of high-confidence boxes that do not overlap, the NMS algorithm must compare every box against every other box. This results in a quadratic ($O(N^2)$) explosion in computational complexity.

Crafting the Phantom Maze

The core mechanism of a sponge attack is the generation of "phantom" or "ghost" objects. Researchers like Derui Wang in the Daedalus paper demonstrate that by optimizing adversarial perturbations, one can flood a detector with a "maze" of redundant, tiny bounding boxes.

To be effective, these phantom objects must satisfy three conditions:

  1. High Confidence: They must pass the initial confidence thresholding so they aren't discarded immediately.
  2. Low Overlap: They must have minimal spatial overlap to ensure the NMS algorithm cannot prune them early in the process.
  3. Stealth: Ideally, they should be placed in the background or be small enough that they don't interfere with the detection of real objects, keeping the attack hidden from simple accuracy checks.

Erh-Chung Chen and Pin-Yu Chen showed in Overload that this strategy can increase inference time on edge devices by over 10x. In safety-critical applications like autonomous driving, this latency spike is catastrophic; a vehicle that takes 500ms to process a frame instead of 30ms is effectively blind to the immediate environment.

Expanding the Attack Surface

The vulnerability isn't limited to single-image detection. Recent work has extended the "sponge" concept to more complex perception tasks:

  • Multi-Object Tracking (MOT): In SlowTrack, Chen Ma and Qi Alfred Chen exploit the logic used to handle temporary occlusions. By re-triggering detections for non-existent objects, they force the system to maintain a massive pool of active tracking threads, leading to exponential increases in data association overhead.
  • Cooperative Perception (CP): Chenyi Wang introduced CP-FREEZER, which targets vehicle-to-vehicle (V2V) communication. By injecting adversarial noise into shared messages, an attacker can "freeze" the decision-making pipeline of an entire fleet of connected vehicles.
  • Hardware-Level Exploits: Manojna Sistla demonstrated that even bit-flips in hardware can be used to trigger these latency spikes, showing that the vulnerability exists at multiple layers of the stack.

Defenses and Reality Checks

Defending against sponge attacks is difficult because standard solutions—like imposing a hard time limit on NMS—often result in the loss of legitimate object detections, creating a different safety risk.

Tianyi Wang proposed a defense called Underload, which uses hardware-adaptive, background-attentive adversarial training. By training the model to be specifically robust against background "phantom" proposals, they were able to restore processing speeds on edge devices like the NVIDIA Jetson from 13 FPS back to 43 FPS.

However, the practical severity of these attacks is still debated. In the EVADE framework, Jean-Philippe Monteuuis and Jonathan Petit argue that many latency attacks fail in realistic deployment scenarios, such as on high-end GPUs or quantized models, where the parallel nature of modern hardware can sometimes absorb the extra computational load. This suggests that while the algorithmic vulnerability is real, its impact is highly dependent on the specific hardware and software implementation of the perception pipeline.

Papers in this version

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 9 papers
  2. Version 1 · 2026-10-04 · 21 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_post_processing_bottleneck_exploitation_80128aa1,
  title = {post-processing bottleneck exploitation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/post-processing-bottleneck-exploitation/rdr_80128aa119a946b4a6b52c17fc6a5281},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_adversarial_sponge_attacks_on_object_det_80128aa1_v1,
  title = {Adversarial Sponge Attacks on Object Detection Pipelines},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/adversarial-sponge-attacks-on-object-detection-pipelines/rdr_80128aa119a946b4a6b52c17fc6a5281/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}