This mug looks like a mug to a human, but through the magic (math) of adversarial machine learning will appear to be a dog to any image classification system. Buy one and start seeing ads online selling dog food and dog entertainment devices. Perfect for bring your dog to work day. Bad at playing fetch.
Adversarial patches don't always transfer well between models. This patch was specifically designed to be effective against YOLOv3.
• Ceramic
• 11 oz mug dimensions: 3.8″ (9.6 cm) in height, 3.2″ (8.2 cm) in diameter
• 15 oz mug dimensions: 4.7″ (11.9 cm) in height, 3.3″ (8.5 cm) in diameter
• 20 oz mug dimensions: 4.3″ (10.9 cm) in height, 3.7″ (9.3 cm) in diameter
• Dishwasher and microwave safe
• Blank product sourced from China
This adversarial designs mug uses a targeted adversarial patch that maximizes the class probability for dog in the bounding box for the mug. The patch is trained with dynamic spline warping to account for the curvature of the mug's surface. You can read the original adversarial patch paper at https://arxiv.org/abs/1712.09665, and learn about the warping method pioneered on adversarial t-shirts at https://arxiv.org/abs/1910.11099.
Adversarial patches don't always transfer well between models. This patch was specifically designed to be effective against YOLOv3.
• Ceramic
• 11 oz mug dimensions: 3.8″ (9.6 cm) in height, 3.2″ (8.2 cm) in diameter
• 15 oz mug dimensions: 4.7″ (11.9 cm) in height, 3.3″ (8.5 cm) in diameter
• 20 oz mug dimensions: 4.3″ (10.9 cm) in height, 3.7″ (9.3 cm) in diameter
• Dishwasher and microwave safe
• Blank product sourced from China
This adversarial designs mug uses a targeted adversarial patch that maximizes the class probability for dog in the bounding box for the mug. The patch is trained with dynamic spline warping to account for the curvature of the mug's surface. You can read the original adversarial patch paper at https://arxiv.org/abs/1712.09665, and learn about the warping method pioneered on adversarial t-shirts at https://arxiv.org/abs/1910.11099.