University of MaineElectrical and Computer Engineering

DASH tamper the feed between the drone and the detector, and a car disappears

A Meta-Attack Framework for Effective and Stealthy Adversarial ExamplesCVPRCORE RANK A*
DASH scorecard
Two things have to hold at once. The detector has to miss the car, and the watch officer pulling up the same frame has to see nothing wrong with it.
Frame as captured
Frame as received
Stages chained in the search
0each stage re-blends the pool and hands the design on
Detector fooled
averaged
Frame untouched
SSIM
Attack in use
DASH
Learns how to blend ten classical attacks, then chains that blend across three stages.
Before the filter
After the filter
hardened classifier · filter shown is neural purification
From the paper: three datasets, seven hardened models, four post-processing defenses, ten base attacks in the pool
99.8%
of frames called the way the attacker wanted, averaged over five defense settings. The best classical attack manages 84%, the best perceptual attack 79%.
94.4
structural similarity to the clean frame, out of 100. The strongest perceptual attack in the comparison sits at 83 while succeeding less often.
52 → 100
average success going from one stage to two. Almost all of the gain comes from chaining, not from any single clever attack.
Results are on RobustBench adversarially trained models with 1,000 sampled test images per dataset. The threat model shown here matches the paper's: an adversary who can edit the frame in transit and who has access to the model, which is the white-box setting the paper assumes throughout. Three caveats worth carrying: baselines were tuned per setting and their best result reported while DASH used one fixed configuration, transfer to a model the weights were not trained on is much weaker than the headline number, and the per-attack weight bars drawn here are illustrative rather than the paper's measured distributions. Aerial photographs by Benjamin Cheng and Iain on Unsplash, used under the Unsplash License.
Funded by the National Science Foundation
For more details, contact
Abdullah Al Nomaan Nafi, lead student researcher abdullah.nafi@maine.edu·Prabuddha Chakraborty, SIEGE Lab PI prabuddha@maine.edu