Ghost Font Bypass and Hardening

A developer has published a bypass for the Ghost Font protection mechanism on Mixfont, using a non-machine-learning attack that achieves a 0.9077 IoU mask overlap and reads 'GHOST'. The same attack against a hardened version yields only 0.0761 IoU and returns 'W4 WZWJ W', demonstrating the effectiveness of the hardening. The attack and defense code are available in the repository, with browser and Python decoders provided.

| Before | After | |---|---| I built this after discovering a way to bypass the protection mechanism by identifying, aligning and removing the moving black elements. The attack uses nine frames, a small vertical-shift search, frame subtraction, thresholding and connected component analysis. It does not use machine learning. Open the Ghost Font page https://www.mixfont.com/ghost-font https://www.mixfont.com/ghost-font , open the developer console, paste code/ghost font browser decoder.js and press Enter. The script captures nine frames from the visible canvas and opens the recovered mask in an overlay. Browser origin rules still apply. If the canvas cannot be read/download or record the animation and use the Python decoder. Python 3.12 and a system ffmpeg executable are required. python -m pip install -r requirements.txt python code/vulnerable ghost font demo.py \ --output attack/ghost-vulnerable.mp4 \ --report attack/ghost-attack.json \ --artifacts attack/artifacts python code/recover ghost font.py \ attack/ghost-vulnerable.mp4 \ --output attack/recovered-red.png \ --classical-ocr \ --ocr-output attack/recovered.txt The measured mask overlap is 0.9077 IoU. The template readout is GHOST . python code/hardened ghost font demo.py \ --output defense/ghost-hardened.mp4 \ --report defense/ghost-defense.json \ --artifacts defense/artifacts The same attack reaches 0.0761 IoU and returns W4 WZWJ W . This result is limited to the global vertical-registration attack implemented here. python code/build attack figure.py --results attack --output figures python code/build defense figure.py --results defense --output figures Both figures are built from the videos and decoder outputs in this repository.