University of MaineElectrical and Computer Engineering

DISARM the drone's AI leaks through its own clock, so fix it on the board it flies

Target Electronic Device Informed Mitigation of Software Runtime Side-Channel VulnerabilitiesIEEE TIFS · 2026
DISARM scorecard
A timing leak is only a leak if the attacker's clock is fine enough to see it. That makes the fix a question about the board and the adversary, not about the source code alone.
Remote clock10 CYC
Physical clock5 CYC
Leak left on the board
cycle gap between the two secret-dependent paths
Gap measured
on the Jetson
What the clock
can resolve
Fix in use
No fix
The routine ships as written. Whether it leaks depends entirely on the board and the attacker.
Runtime added
Lines added
cost bars: Java benchmark set, Jetson Nano · runtime scale 0 to 200 percent
From the paper: 22 benchmarks, C/C++ and Java, measured on five real embedded boards
up to 93%
cut in the timing gap between secret-dependent paths, measured on the device itself, driving the leak below the attacker's resolution instead of below zero.
≈ 40%
faster than PENDULUM's repaired code and 6 to 10 percent faster than DifFuzzAR's, while adding fewer lines than either.
22 / 5
benchmarks across five boards. Every DISARM repair compiled, passed regression tests, and met its threshold. Two of the ten Java baselines did not.
What this demo does and does not show. The severity numbers are DISARM's own cycle measurements on a Jetson Nano running a convolution benchmark, not an end-to-end secret recovery: the paper measures the gap an attacker would have to resolve, it does not run the attack. "Secure" here means the gap sits below an assumed attacker resolution on that board, which is weaker than a constant-time proof. The head-to-head cost comparison is on the Java benchmarks only, because PENDULUM and DifFuzzAR do not natively support C or C++. The paper also notes that its padding uses a fixed pattern, which may itself be distinguishable under power, electromagnetic, or fault-based observation.
Funded by the U.S. Army Engineer Research and Development Center
For more details, contact
Tasneem Suha, lead student researcher tasneem.suha@maine.edu·Prabuddha Chakraborty, SIEGE Lab PI prabuddha@maine.edu