It was a noble experiment. I wanted to see if I could build my own synthetic persona from scratch, without relying on a more plug-and-play cloud API such as the one David Maimon showed me this summer at Georgia State University.
My motivation? I hoped to safely (and ethically) engage with deep-faked scammers through my own deep-faked identity. I opted first to learn the hard way.
The Local Grind: Air-Gaps & 26 hours of training
I deployed a powerful gaming laptop running a graphics card with what I thought was sufficient VRAM (12 GB). As a makeshift “air gapped” machine that still required an internet connection with a light footprint, I purchased a SIM card with cash, and loaded it onto a wiped Android mobile device. This would serve as the tethered internet connection for the laptop. My intent was to avoid detection (or being tracked by fraudsters) through my regular IP address and daily devices.
Guided by a compliant LLM, I downloaded Fooocus, an open-source image generator hosted on GitHub. I took a selfie, then had the software generate variations of my face in different angles, facial expressions, and even eyes closed.
Next, I recorded a 3-minute video as I spoke to camera, and fed it, along with my synthetic selfies from Fooocus into DeepFaceLab, another open-source generator.
It was a grind for both me and my computer. I left the neural network to train for 26 hours. When it was time to composite the final video, I had to manually dial in the interactive merger (such as fading and blurring the synthetic mask’s digital seams), and then command the GPU to batch-render 5,597 frames.
The result was…flawed. As the software processed everything frame-by-frame, I ran into temporal jitter (how Star Trek). The mask trembled slightly along my jawline, and the teeth and lip movements lacked natural fluidity.
The Cloud Pivot: Plug-and-Play Reality
Maimon reminded me that modern fraud is no longer a series of isolated schemes, but a specialized criminal infrastructure. He showed me what was possible with Decart’s Lucy.AI model. That cloud software uses a self-anchoring mechanism that builds consistency directly from its own output history. Instead of waiting for 26 hours on a gaming rig, I can use any computer, write a simple Python script, and send the video to a cloud API. It processes at 1080p, 30 frames per second, and yields perfectly consistent, real-time time digital disguises without trembling masks. I had originally aimed for a local model as I thought it would be more stable in a live Zoom or Teams interaction tethered to a mobile hotspot. But what I created on my own was clearly not going to pass muster with any scammer.
Here’s a short clip of Maimon easily masquerading as me via a simple headshot on the web and Lucy.AI:
Defending the Human Perimeter
The barrier to entry for highly convincing synthetic video clearly doesn’t require a $3,000 setup. An API key is more than sufficient, and that’s what available to the world. Still, I appreciated getting my hands “dirty” in the machinery of synthetic media. While I agonized over all the distinct points of failure when configuring my tools, I now recognize how cloud tools and an asynchronous API call that costs pennies per minute is a specialized parlor trick upturned into an industrialized global threat.
Yet, even as we admit that seeing is no longer believing, there is hope. The deepfakes do leave behind algorithmic footprints that are detectable by “good guy” software engines: the absence of micro-latencies and even the acoustic reflections of a physical room, along with the presence of digital clipping and unnatural uniform frequencies.
And, as I learned in trying to create an alternative digital footprint, it’s not so easy to prompt years of genuine behavioral history into existence. Yes, you can easily purchase a social security number or utility bill on the dark web. Yes, you can create a fake LinkedIn persona with fictional experience and education. But, our messy digital exhaust of online behaviors, relationships, and provenance are much harder to replicate. Who will accept your connection invites after all as a synthetic persona? What are the signals that Big Tech’s algorithms are constantly scanning for as part of that “Reverse Turing Test?” That demands proof of the timeline behind the mask. Context matters more than ever, and we should not hesitate to ask more questions to fill in those blanks.