why use negative no_w_metric and w_metric?

tree-ring-watermark/run_tree_ring_watermark.py

Line 141 in 3015283

no_w_metrics.append(-no_w_metric)

tree-ring-watermark/run_tree_ring_watermark.py

Line 142 in 3015283

w_metrics.append(-w_metric)

Why do you use negative no_w_metric and w_metric instead of the postive ones?

Hi, @liwd190019, thanks for asking. It's a good question. For preds that we pass to metrics.roc_curve is the confidence score of being watermarked. If it's a watermarked image, this score should be larger. We negate it, since our metric is the distance, and we know that the distance is smaller for the watermarked images.

Let me know if you have further questions!