Evaluating Google’s reCAPTCHA v2 Security Through Single-Category Object Detection

Abstract

Google’s reCAPTCHA v2 relies on image-based object recognition challenges to distinguish humans from automated bots, but advances in AI object detection increasingly threaten its security assumptions. This thesis evaluates reCAPTCHA v2’s resistance to attack by fine-tuning a YOLOv9 model exclusively on cars, using 8,638 training images, to overcome the catastrophic forgetting problem identified in prior work. A novel dataset of 1,132 annotated reCAPTCHA v2 3×3 car challenge images was collected directly from Google’s system to enable domain-specific evaluation. The fine-tuned model achieved an F1 Score of 0.72 on a generic car control dataset and 0.718 on the reCAPTCHA dataset, but this narrow gap obscures a critical divergence: Precision rose to 96.8% while Recall fell to 57.0% on the reCAPTCHA dataset. This reveals that reCAPTCHA’s anti-recognition techniques primarily suppress detections rather than cause misclassifications, and since solving a challenge requires selecting every car-containing tile, a 43% miss rate would lead to high failure rates in practice. The image-based security of reCAPTCHA v2 is real but not absolute, and its defensive value lies in suppressing detections rather than generating false ones.