logseq/pages/hls__icse2023-poster-paper6_1676762920568_0.md

file:: icse2023-poster-paper6_1676762920568_0.pdf file-path:: ../assets/icse2023-poster-paper6_1676762920568_0.pdf

• DISTROFAIR automatically learns the distribution (e.g., number/orientation) of objects in a set of images and systematically mutates objects in the images to become OOD using three semantic-preserving image mutations – object deletion, ls-type:: annotation hl-page:: 1 hl-color:: green id:: 63f16015-e1e1-4eb5-94c7-a43da9c53d96
• class-level fairness violations ls-type:: annotation hl-page:: 1 hl-color:: yellow id:: 63f1cbbc-4a1a-491e-8465-8b576e2c949d
• image recognition ls-type:: annotation hl-page:: 1 hl-color:: green id:: 63f1cbf5-4d2f-474f-90fc-a16954ce93b3
• systematic testing of image classification systems, to detect potential bias against certain classes, is of critical importance. ls-type:: annotation hl-page:: 1 hl-color:: purple id:: 63f1cc0d-033b-4d93-9a40-528da681b3c8
• DISTROFAIR) identifies the classes that are subject to unfair treatment (i.e., unusually high error rates) ls-type:: annotation hl-page:: 1 hl-color:: green id:: 63f1cc27-3991-4f6b-b669-631f7588433c
• DISTROFAIR learns the distribution of objects detected in a set of images and systematically generates a set of images, named OOD images, outside such a distributio ls-type:: annotation hl-page:: 1 hl-color:: green id:: 63f1cc88-1792-40c0-b9d7-c7e83b282382
• OOD images are generated by leveraging this information and using semantic-preserving mutation operators(e.g., insertion, deletion and rotation of objects). ls-type:: annotation hl-page:: 1 hl-color:: yellow id:: 63f1cfa9-832f-49b9-b6b2-13d64af880a5
• OOD images are about 80% as realistic as the original images. ls-type:: annotation hl-page:: 1 hl-color:: yellow id:: 63f1d03d-00a5-4dcc-81e9-361e5f340323
• 12K erroneous OOD image ls-type:: annotation hl-page:: 1 hl-color:: yellow id:: 63f1d062-fcb8-41c9-9207-95dfb9626465
• discovery of fairness errors ls-type:: annotation hl-page:: 1 hl-color:: green id:: 63f1d089-d192-4733-bce9-b0ad83884753
• lasses exhibiting an error rate higher than mean error rate across all classes would then be tagged as being unfairly treated. ls-type:: annotation hl-page:: 2 hl-color:: green id:: 63f1d110-b8cd-475f-9e04-8b204ac2c7cd
• Classes exhibiting an error rate higher than mean error rate across all classes would then be tagged as being unfairly treated ls-type:: annotation hl-page:: 2 hl-color:: green id:: 63f1d113-6008-4087-8c2c-a6b7027135d8
• lasses exhibiting an error rate higher than mean error rate across all classes would then be tagged as being unfairly treated. ls-type:: annotation hl-page:: 2 hl-color:: green id:: 63f1d116-a003-43fe-883d-cd87098c8183
• DISTROFAIR induce class-level group fairness violation. ls-type:: annotation hl-page:: 2 hl-color:: green id:: 63f1d13d-100d-4d73-8454-204ba1ce3d0a
• OOD image mutation ls-type:: annotation hl-page:: 2 hl-color:: green id:: 63f1d146-8587-4e74-b403-5f0cedc9a67d
• insertion mutation operation ls-type:: annotation hl-page:: 2 hl-color:: green id:: 63f1d17d-2b45-44b5-82f0-897d300f06a9
• developer is more than two times likely (up to131%) to find class-level fairness errors with OOD mutations than ID. ls-type:: annotation hl-page:: 2 hl-color:: green id:: 63f1d1a1-90e3-4f56-acfb-41d12664ca3d
• ystematic approach to discover class-level fairness violations in image classification task ls-type:: annotation hl-page:: 2 hl-color:: green id:: 63f1d1dc-4632-42b7-b909-19f054d73cde
• emantic preserving mutation operation ls-type:: annotation hl-page:: 2 hl-color:: yellow id:: 63f1d1ec-bc72-4ae3-b565-e46e612e9dce