⚡ electrical engineering

TriSweep: A Four-Drone Swarm Framework for Electromagnetic Side-Channel Analysis

TriSweep is a simulation framework proposing a four-drone swarm architecture that utilizes spatially specialized collectors and coherent signal combining to successfully perform autonomous, standoff electromagnetic side-channel analysis on masked microcontrollers at distances of 0.25–1.5 meters, achieving significant key recovery improvements over single-drone approaches despite the lack of physical hardware implementation.

Original authors: Eric Yocam, Varghese Vaidyan

Published 2026-05-22

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Eric Yocam, Varghese Vaidyan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to listen to a secret conversation happening inside a locked room. Usually, to hear the whispers, you have to press your ear right up against the door. But what if you could stand across the street and still hear every word?

That is the idea behind TriSweep, a new computer simulation created by researchers Eric Yocam and Varghese Vaidyan. They aren't building a real flying robot yet; instead, they built a "virtual playground" to test if a team of four tiny drones could steal secret encryption keys from a computer chip from a distance, without ever touching it.

Here is how their "drone squad" works, explained simply:

The Problem: The "Whisper" is Too Faint

When a computer chip (like the one in a smart card or a router) does math to encrypt data, it leaks tiny bits of electromagnetic energy. Think of this like a faint whisper.

The Old Way: Hackers usually have to put a sensor millimeters away from the chip to hear the whisper clearly.
The New Threat: The researchers asked, "What if a hacker flies a drone close to a window or over a fence?" They found that current security assumes the hacker can't get close enough. TriSweep tests if they actually can.

The Solution: A Four-Drone Team

Instead of one drone trying to do everything, TriSweep uses a team of four drones with specific jobs, working together like a well-rehearsed orchestra.

Drone A (The Anchor): This drone stays still and acts as the "conductor." It keeps time and makes sure everyone is on the same page.
Drone B (The Mask Listener): The computer chip is wearing a "mask" (a security trick that hides the secret). Drone B is specialized to listen for the moment the chip puts the mask on.
Drone C (The Cipher Listener): This drone listens for the moment the chip does the actual math (the secret part) while wearing the mask.
Drone D (The Accumulator): This is the "brain." It sits a bit further back, receives the recordings from the other three, and does the heavy lifting.

The Magic Trick: "Noise Canceling" the Secret

Here is the clever part. The chip's "mask" is like static noise that hides the secret message.

Drone B hears the "mask noise."
Drone C hears the "secret message + mask noise."
Drone D takes both recordings and multiplies them together in a special way.

Imagine you have two recordings of a song. One has a loud cough in it, and the other has the same cough but no music. If you mix them perfectly, the cough cancels itself out, and suddenly, the music becomes clear. Drone D does this mathematically to cancel out the "mask" and reveal the secret key.

The Results: How Well Did It Work?

Since they haven't built the real drones yet, they ran this entire scenario on a super-fast computer using real data from other security tests.

The Distance: They tested distances from 10 inches (0.25 meters) to about 5 feet (1.5 meters).
The Teamwork: One drone alone was terrible at guessing the key (it was like trying to find a needle in a haystack). But when the three listening drones worked together, they got 4.8 decibels louder (like turning up the volume significantly).
The Success: With the full four-drone team, the simulation successfully guessed the secret key with a very high confidence level. It was about 10 times better than a single drone could do.
The Jitter Fix: Real drones shake a little when they hover (like a shaky hand holding a camera). The researchers added a "stabilizer" in their code that lined up the recordings perfectly, fixing the shaking problem and making the attack even more accurate.

The Catch (What the Paper Actually Says)

It is important to know what this paper does not say:

No Real Drones: They did not build or fly any physical drones. Everything happened inside a computer simulation.
No Real Hardware: They didn't build the special antennas or the "brain" drone yet.
The Next Step: The researchers say the very next step is to actually build the prototype and fly it to see if the real world behaves like their computer model.

The Bottom Line

TriSweep is a "proof of concept" simulation. It shows that if you have four drones working together as a team, you might be able to steal secret codes from a distance by listening to the electromagnetic "whispers" of a computer chip. It suggests that hiding a device behind a locked door might not be enough if a drone can hover nearby and listen in.

Technical Summary: TriSweep – A Four-Drone Swarm Framework for Electromagnetic Side-Channel Analysis

Problem Statement
Electromagnetic (EM) side-channel analysis (SCA) traditionally relies on a threat model where an adversary places a probe at millimeter-to-centimeter distances from a target device. This assumption renders physical security measures, such as locked panels and enclosures, effective. However, the proliferation of commercial off-the-shelf (COTS) unmanned aerial vehicles (UAVs) and low-cost software-defined radio (SDR) hardware challenges this model. An adversary capable of hovering a drone at standoff distances of 0.25–1.5 meters can bypass proximity-based physical security without direct access. Furthermore, modern embedded implementations often employ first-order masking schemes to randomize intermediate values, requiring second-order analysis to recover keys. Traditional second-order attacks must algorithmically separate leakage events (e.g., mask loading and masked computation) from a single trace, a process sensitive to timing jitter and noise. The paper posits that a multi-drone swarm could spatially separate these leakage events, physically canceling the mask without algorithmic preprocessing.

Methodology
TriSweep is a simulation framework designed to evaluate a four-drone swarm architecture for autonomous standoff EM-SCA. The system does not utilize fabricated hardware; instead, it combines real, publicly available ANSSI ASCAD EM datasets (ATmega8515 masked AES-128) with a physics-based free-space path-loss noise model.

The framework consists of four distinct drone roles:

Drone A (Anchor): Hovers at a reference distance ( $D_{ref} = 0.25$ m) to provide a clock reference and full-spectrum EM collection.
Drone B (Mask Probe): Specialized to capture the mask-register loading leakage window.
Drone C (Cipher Probe): Specialized to capture the masked SubBytes output leakage window.
Drone D (Accumulator): A stationary unit that receives synchronized IQ traces from the other three drones.

The attack pipeline relies on seven algorithms:

Communication & Targeting: A 50 Hz heartbeat protocol coordinates capture triggers and target localization via Time Difference of Arrival (TDOA).
Swarm Repositioning: Drones B and C autonomously reposition every 200 ms to maximize distributed Fisher information.
Clock Synchronization: A two-stage protocol (GPSDO coarse alignment followed by 1 kHz pilot-tone cross-correlation) achieves sub-10 ns inter-drone jitter, enabling coherent combining.
Coherent Combining: Drone D performs maximum ratio combining (MRC) of the synchronized traces, yielding a theoretical 4.8 dB SNR gain ( $10 \log_{10}(3)$ ).
Second-Order Mask Cancellation: Drone D computes the centered product of the traces from Drone B and Drone C ( $X_{SO} = (T_B - \bar{T}_B) \cdot (T_C - \bar{T}_C)$ ). This physical spatial separation allows the mask $r$ to cancel out without knowledge of its value.
Key-Rank Accumulation: The system aggregates log-likelihoods from first-order traces, the second-order product, and an optional two-channel Convolutional Neural Network (CNN).

Key Contributions

Four-Drone Architecture: A novel design separating leakage sources across dedicated collector drones (Anchor, Mask Probe, Cipher Probe) and a central Accumulator.
Physical Second-Order Cancellation: The first proposed framework to achieve second-order mask cancellation via spatial decomposition and coherent combining, rather than purely algorithmic separation of a single trace.
Swarm Consensus & Synchronization: Protocols for distributed Fisher information maximization and sub-10 ns clock synchronization tailored for airborne platforms.
Profiling-Trace Alignment: A cross-correlation alignment method that compensates for drone hover vibration (simulated as 100-sample jitter), reducing key rank significantly on desynchronized datasets.
Simulation Framework: A validated simulation environment using real ASCAD datasets and calibrated noise models to predict the performance of a physical system before fabrication.

Results
Evaluated against three real ASCAD datasets (primary masked, 50-sample desync, and 100-sample desync), the simulation yielded the following:

Key Rank Improvement: On the primary masked dataset at 0.25 m, the four-drone system achieved a simulated key rank of 18 ± 1.7 (five-seed validation), a tenfold improvement over the single-drone baseline rank of 197.
SNR Gain: Coherent combining of three collectors provided a 4.8 dB SNR gain, consistent with theoretical predictions.
Jitter Compensation: Profiling-trace cross-correlation alignment reduced the key rank on the 100-sample jitter dataset from 89 (single-drone, unaligned) to 21.
CNN Performance: A two-channel CNN in the Accumulator converged to a loss of 0.454 (vs. random baseline 5.545) and improved ranks on desynchronized datasets, though it overfit on the clean masked data compared to the manual second-order product.
Distance Sensitivity: The system maintained a key rank of 20 at 0.25 m and 25 at 1.5 m on the primary dataset, demonstrating that second-order combining compensates for SNR loss at greater standoff distances.

Significance and Claims
The paper claims that TriSweep establishes a "qualitatively new threat vector" by combining aerial mobility, multi-receiver coherent gain, and autonomous repositioning to defeat masked cryptographic implementations at standoff distances. The primary significance lies in the physical spatial decomposition of leakage events, which allows for mask cancellation without algorithmic preprocessing or knowledge of the mask value.

The authors explicitly state that no physical hardware has been fabricated; all results are derived from simulation using real datasets and a noise model. Consequently, the paper positions TriSweep as a design specification and quantitative prediction tool. The authors acknowledge that physical deployment faces unmodeled challenges, such as propeller electromagnetic interference (EMI) and vibration-induced phase noise, which may degrade the simulated SNR. The immediate next step identified is the construction and flight of a physical prototype to validate the framework's predictions and quantify the gap between simulation and reality.