In Progress

Adversarial ML & Surveillance Commentary

Signal
Jammer

A live pipeline that poisons your image before a surveillance model can read it. Speculative armor against machine vision.

Type Software Installation
Year 2025
Output CRT Monitor (planned)
Stack TouchDesigner, StreamDiffusion, GLSL
Status In Progress
The Idea

Who Is the Camera For?

Signal Jammer began in a graduate research context, a deep dive into adversarial machine learning for an AI for Artists course at Pratt Institute. The underlying question was practical and political at once: if AI systems can be deceived by imperceptible noise, what does that mean for the people those systems are used against?

Surveillance infrastructure powered by computer vision is now embedded in public space at scale. Facial recognition, gait analysis, behavioral prediction models. These systems are deployed by governments and corporations with limited accountability, and they disproportionately harm the people with the least power to contest them. They profile. They flag. They kill.

Signal Jammer is a speculative tool. It asks: what would it look like if you could make yourself illegible to the machine, in real time, using the machine's own weaknesses against it?

The piece does not offer a solution. It makes the problem visible. Standing in front of a CRT monitor watching your own image dissolve into surveillance glitch and digital decay, you understand something about what it means to be seen by a system that was never designed to see you as a person.

Pipeline

How It Works

The system runs as a continuous real-time loop inside TouchDesigner. A live webcam feed passes through two layers of adversarial processing before compositing, producing an output that is simultaneously a portrait and a document of machine failure.

01
Live Capture
Webcam input piped into TouchDesigner as a live video stream. The low resolution is intentional — it matches the aesthetic register of surveillance infrastructure.
TouchDesigner
02
GLSL Adversarial Noise
A custom GLSL shader applies adversarial perturbation noise to the live feed. The noise is structured to disrupt machine vision systems while remaining visually present — unlike traditional FGSM approaches, the interference is intentionally visible as a formal element of the piece.
GLSL
03
StreamDiffusion
The perturbed feed passes through StreamDiffusion TOX, a real-time image-to-image diffusion pipeline running inside TouchDesigner. The model interprets the adversarial noise as instruction, generating a continuous stream of algorithmically degraded output from the live input.
StreamDiffusion TOX
04
Composite
The clean source video feeds through edge detection to produce a live outline — a white ghost of the original figure composited over the diffusion output. A second pass of GLSL adversarial noise is applied to the composite for additional visual disruption. The result layers three states of the image simultaneously: the original, its machine interpretation, and its outline.
TouchDesigner
05
CRT Output
Planned output to a CRT monitor. The CRT's natural scan lines, phosphor glow, and color bleed will interact with the composited image, adding a final physical layer of degradation to the algorithmically corrupted feed.
CRT Monitor
The Object

What You See

The CRT monitor is not incidental. It is the piece's primary material choice. Surveillance cameras and the systems built around them have been operating since before the flat screen, and the CRT carries that history in its body. Its scan lines, its hum, its curved glass, its slightly wrong colors.

What appears on the screen is three versions of you at once. The diffusion model's interpretation of your perturbed image — smeared, decayed, alien. Over that, a white ghost: the clean edge of your original figure, traced live and composited on top. And over everything, a second layer of adversarial noise that refuses to let the image settle.

The ghost is the most unsettling part. It is precise where the diffusion output is not. It says: I know where you are. I just refuse to show you clearly.

Environment
TouchDesigner
Diffusion
StreamDiffusion TOX
Noise
GLSL Shader
Composite
Edge Detection + Ghost Overlay
Input
Live Webcam
Output
CRT Monitor (planned)

Back to

All Work

↗