ML Connoisseur

uoftctf2026

Task: Obfuscated PyTorch digit classifier with a hidden backdoor that returns class 10 when input features match a secret reference tensor (MSE < 0.001). Solution: Performed a model inversion attack using multi-phase gradient descent optimization with a Straight-Through Estimator for quantization-aware fine-tuning to generate a PNG image whose features match the hidden reference.