a theoretical looks at adversarial examples
play

A THEORETICAL LOOKS AT ADVERSARIAL EXAMPLES Tom Goldstein and also - PowerPoint PPT Presentation

A THEORETICAL LOOKS AT ADVERSARIAL EXAMPLES Tom Goldstein and also Ali Shafahi, Ronny Huang, Mahyar Najibi, Octavian Suciu, Christoph Studer, Soheil Feizi, Tudor Dumitras OVERVIEW Why is optimization so easy on neural nets? What are


  1. A THEORETICAL LOOKS AT ADVERSARIAL EXAMPLES Tom Goldstein …and also… Ali Shafahi, Ronny Huang, Mahyar Najibi, Octavian Suciu, Christoph Studer, Soheil Feizi, Tudor Dumitras

  2. OVERVIEW Why is optimization so easy on neural nets? What are adversarial examples, and what are their risks? Poison attacks Are they an escapable problem?

  3. <latexit sha1_base64="d96mskUPBfSUp9ih23g0ari3WA=">ACAXicbVA9SwNBEJ2LX/H8ilraLAbBKtyJoGXQxjKiSYQkhL3NXLJkb+/Y3VPCkcraVn+Dndj6S/wJ/gs3yRUm8cHA470ZuYFieDaeN63U1hZXVvfKG6W9s7u3ul/YOGjlPFsM5iEauHgGoUXGLdcCPwIVFIo0BgMxheT/zmIyrNY3lvRgl2ItqXPOSMGivdPXV5t1T2Kt4UZJn4OSlDjlq39NPuxSyNUBomqNYt30tMJ6PKcCZw7LZTjQlQ9rHlqWSRqg72fTUMTmxSo+EsbIlDZmqfycyGmk9igLbGVEz0IveRPzPa6UmvOxkXCapQclmi8JUEBOTyd+kxUyI0aWUKa4vZWwAVWUGZuO687tkZxhaK2xzcZfTGKZNM4qvlfxb8/L1as8pSIcwTGcg8XUIUbqEdGPThBV7hzXl23p0P53PWnDymUOYg/P1C3FGlx0=</latexit> <latexit sha1_base64="d96mskUPBfSUp9ih23g0ari3WA=">ACAXicbVA9SwNBEJ2LX/H8ilraLAbBKtyJoGXQxjKiSYQkhL3NXLJkb+/Y3VPCkcraVn+Dndj6S/wJ/gs3yRUm8cHA470ZuYFieDaeN63U1hZXVvfKG6W9s7u3ul/YOGjlPFsM5iEauHgGoUXGLdcCPwIVFIo0BgMxheT/zmIyrNY3lvRgl2ItqXPOSMGivdPXV5t1T2Kt4UZJn4OSlDjlq39NPuxSyNUBomqNYt30tMJ6PKcCZw7LZTjQlQ9rHlqWSRqg72fTUMTmxSo+EsbIlDZmqfycyGmk9igLbGVEz0IveRPzPa6UmvOxkXCapQclmi8JUEBOTyd+kxUyI0aWUKa4vZWwAVWUGZuO687tkZxhaK2xzcZfTGKZNM4qvlfxb8/L1as8pSIcwTGcg8XUIUbqEdGPThBV7hzXl23p0P53PWnDymUOYg/P1C3FGlx0=</latexit> <latexit sha1_base64="d96mskUPBfSUp9ih23g0ari3WA=">ACAXicbVA9SwNBEJ2LX/H8ilraLAbBKtyJoGXQxjKiSYQkhL3NXLJkb+/Y3VPCkcraVn+Dndj6S/wJ/gs3yRUm8cHA470ZuYFieDaeN63U1hZXVvfKG6W9s7u3ul/YOGjlPFsM5iEauHgGoUXGLdcCPwIVFIo0BgMxheT/zmIyrNY3lvRgl2ItqXPOSMGivdPXV5t1T2Kt4UZJn4OSlDjlq39NPuxSyNUBomqNYt30tMJ6PKcCZw7LZTjQlQ9rHlqWSRqg72fTUMTmxSo+EsbIlDZmqfycyGmk9igLbGVEz0IveRPzPa6UmvOxkXCapQclmi8JUEBOTyd+kxUyI0aWUKa4vZWwAVWUGZuO687tkZxhaK2xzcZfTGKZNM4qvlfxb8/L1as8pSIcwTGcg8XUIUbqEdGPThBV7hzXl23p0P53PWnDymUOYg/P1C3FGlx0=</latexit> <latexit sha1_base64="d96mskUPBfSUp9ih23g0ari3WA=">ACAXicbVA9SwNBEJ2LX/H8ilraLAbBKtyJoGXQxjKiSYQkhL3NXLJkb+/Y3VPCkcraVn+Dndj6S/wJ/gs3yRUm8cHA470ZuYFieDaeN63U1hZXVvfKG6W9s7u3ul/YOGjlPFsM5iEauHgGoUXGLdcCPwIVFIo0BgMxheT/zmIyrNY3lvRgl2ItqXPOSMGivdPXV5t1T2Kt4UZJn4OSlDjlq39NPuxSyNUBomqNYt30tMJ6PKcCZw7LZTjQlQ9rHlqWSRqg72fTUMTmxSo+EsbIlDZmqfycyGmk9igLbGVEz0IveRPzPa6UmvOxkXCapQclmi8JUEBOTyd+kxUyI0aWUKa4vZWwAVWUGZuO687tkZxhaK2xzcZfTGKZNM4qvlfxb8/L1as8pSIcwTGcg8XUIUbqEdGPThBV7hzXl23p0P53PWnDymUOYg/P1C3FGlx0=</latexit> <latexit sha1_base64="UfjSuN0X3dJRIR9aQfKBZasZn34=">ACAXicbVA9SwNBEJ2LX/H8ilraLAbBKtyJoGXQxjKiSYQkhL3NXLJkb+/Y3VPCkcraVn+Dndj6S/wJ/gs3yRUm8cHA470ZuYFieDaeN63U1hZXVvfKG6W9s7u3ul/YOGjlPFsM5iEauHgGoUXGLdcCPwIVFIo0BgMxheT/zmIyrNY3lvRgl2ItqXPOSMGivdPX9bqnsVbwpyDLxc1KGHLVu6afdi1kaoTRMUK1bvpeYTkaV4Uzg2G2nGhPKhrSPLUsljVB3sumpY3JilR4JY2VLGjJV/05kNJ6FAW2M6JmoBe9ifif10pNeNnJuExSg5LNFoWpICYmk79JjytkRowsoUxeythA6oMzYd153bIznD0Fpjm42/mMQyaZxVfK/i356Xq1d5SkU4gmM4BR8uoAo3UIM6MOjDC7zCm/PsvDsfzuesteDkM4cwB+frFxdeluU=</latexit> <latexit sha1_base64="UfjSuN0X3dJRIR9aQfKBZasZn34=">ACAXicbVA9SwNBEJ2LX/H8ilraLAbBKtyJoGXQxjKiSYQkhL3NXLJkb+/Y3VPCkcraVn+Dndj6S/wJ/gs3yRUm8cHA470ZuYFieDaeN63U1hZXVvfKG6W9s7u3ul/YOGjlPFsM5iEauHgGoUXGLdcCPwIVFIo0BgMxheT/zmIyrNY3lvRgl2ItqXPOSMGivdPX9bqnsVbwpyDLxc1KGHLVu6afdi1kaoTRMUK1bvpeYTkaV4Uzg2G2nGhPKhrSPLUsljVB3sumpY3JilR4JY2VLGjJV/05kNJ6FAW2M6JmoBe9ifif10pNeNnJuExSg5LNFoWpICYmk79JjytkRowsoUxeythA6oMzYd153bIznD0Fpjm42/mMQyaZxVfK/i356Xq1d5SkU4gmM4BR8uoAo3UIM6MOjDC7zCm/PsvDsfzuesteDkM4cwB+frFxdeluU=</latexit> <latexit sha1_base64="UfjSuN0X3dJRIR9aQfKBZasZn34=">ACAXicbVA9SwNBEJ2LX/H8ilraLAbBKtyJoGXQxjKiSYQkhL3NXLJkb+/Y3VPCkcraVn+Dndj6S/wJ/gs3yRUm8cHA470ZuYFieDaeN63U1hZXVvfKG6W9s7u3ul/YOGjlPFsM5iEauHgGoUXGLdcCPwIVFIo0BgMxheT/zmIyrNY3lvRgl2ItqXPOSMGivdPX9bqnsVbwpyDLxc1KGHLVu6afdi1kaoTRMUK1bvpeYTkaV4Uzg2G2nGhPKhrSPLUsljVB3sumpY3JilR4JY2VLGjJV/05kNJ6FAW2M6JmoBe9ifif10pNeNnJuExSg5LNFoWpICYmk79JjytkRowsoUxeythA6oMzYd153bIznD0Fpjm42/mMQyaZxVfK/i356Xq1d5SkU4gmM4BR8uoAo3UIM6MOjDC7zCm/PsvDsfzuesteDkM4cwB+frFxdeluU=</latexit> <latexit sha1_base64="UfjSuN0X3dJRIR9aQfKBZasZn34=">ACAXicbVA9SwNBEJ2LX/H8ilraLAbBKtyJoGXQxjKiSYQkhL3NXLJkb+/Y3VPCkcraVn+Dndj6S/wJ/gs3yRUm8cHA470ZuYFieDaeN63U1hZXVvfKG6W9s7u3ul/YOGjlPFsM5iEauHgGoUXGLdcCPwIVFIo0BgMxheT/zmIyrNY3lvRgl2ItqXPOSMGivdPX9bqnsVbwpyDLxc1KGHLVu6afdi1kaoTRMUK1bvpeYTkaV4Uzg2G2nGhPKhrSPLUsljVB3sumpY3JilR4JY2VLGjJV/05kNJ6FAW2M6JmoBe9ifif10pNeNnJuExSg5LNFoWpICYmk79JjytkRowsoUxeythA6oMzYd153bIznD0Fpjm42/mMQyaZxVfK/i356Xq1d5SkU4gmM4BR8uoAo3UIM6MOjDC7zCm/PsvDsfzuesteDkM4cwB+frFxdeluU=</latexit> CONVOLUTIONAL NET y x w 1 One’s hot label 0 0 1 0 w i ball car hat grumpy cat f ( x ; w ) → y image weights

  4. <latexit sha1_base64="UfjSuN0X3dJRIR9aQfKBZasZn34=">ACAXicbVA9SwNBEJ2LX/H8ilraLAbBKtyJoGXQxjKiSYQkhL3NXLJkb+/Y3VPCkcraVn+Dndj6S/wJ/gs3yRUm8cHA470ZuYFieDaeN63U1hZXVvfKG6W9s7u3ul/YOGjlPFsM5iEauHgGoUXGLdcCPwIVFIo0BgMxheT/zmIyrNY3lvRgl2ItqXPOSMGivdPX9bqnsVbwpyDLxc1KGHLVu6afdi1kaoTRMUK1bvpeYTkaV4Uzg2G2nGhPKhrSPLUsljVB3sumpY3JilR4JY2VLGjJV/05kNJ6FAW2M6JmoBe9ifif10pNeNnJuExSg5LNFoWpICYmk79JjytkRowsoUxeythA6oMzYd153bIznD0Fpjm42/mMQyaZxVfK/i356Xq1d5SkU4gmM4BR8uoAo3UIM6MOjDC7zCm/PsvDsfzuesteDkM4cwB+frFxdeluU=</latexit> <latexit sha1_base64="UfjSuN0X3dJRIR9aQfKBZasZn34=">ACAXicbVA9SwNBEJ2LX/H8ilraLAbBKtyJoGXQxjKiSYQkhL3NXLJkb+/Y3VPCkcraVn+Dndj6S/wJ/gs3yRUm8cHA470ZuYFieDaeN63U1hZXVvfKG6W9s7u3ul/YOGjlPFsM5iEauHgGoUXGLdcCPwIVFIo0BgMxheT/zmIyrNY3lvRgl2ItqXPOSMGivdPX9bqnsVbwpyDLxc1KGHLVu6afdi1kaoTRMUK1bvpeYTkaV4Uzg2G2nGhPKhrSPLUsljVB3sumpY3JilR4JY2VLGjJV/05kNJ6FAW2M6JmoBe9ifif10pNeNnJuExSg5LNFoWpICYmk79JjytkRowsoUxeythA6oMzYd153bIznD0Fpjm42/mMQyaZxVfK/i356Xq1d5SkU4gmM4BR8uoAo3UIM6MOjDC7zCm/PsvDsfzuesteDkM4cwB+frFxdeluU=</latexit> <latexit sha1_base64="UfjSuN0X3dJRIR9aQfKBZasZn34=">ACAXicbVA9SwNBEJ2LX/H8ilraLAbBKtyJoGXQxjKiSYQkhL3NXLJkb+/Y3VPCkcraVn+Dndj6S/wJ/gs3yRUm8cHA470ZuYFieDaeN63U1hZXVvfKG6W9s7u3ul/YOGjlPFsM5iEauHgGoUXGLdcCPwIVFIo0BgMxheT/zmIyrNY3lvRgl2ItqXPOSMGivdPX9bqnsVbwpyDLxc1KGHLVu6afdi1kaoTRMUK1bvpeYTkaV4Uzg2G2nGhPKhrSPLUsljVB3sumpY3JilR4JY2VLGjJV/05kNJ6FAW2M6JmoBe9ifif10pNeNnJuExSg5LNFoWpICYmk79JjytkRowsoUxeythA6oMzYd153bIznD0Fpjm42/mMQyaZxVfK/i356Xq1d5SkU4gmM4BR8uoAo3UIM6MOjDC7zCm/PsvDsfzuesteDkM4cwB+frFxdeluU=</latexit> <latexit sha1_base64="UfjSuN0X3dJRIR9aQfKBZasZn34=">ACAXicbVA9SwNBEJ2LX/H8ilraLAbBKtyJoGXQxjKiSYQkhL3NXLJkb+/Y3VPCkcraVn+Dndj6S/wJ/gs3yRUm8cHA470ZuYFieDaeN63U1hZXVvfKG6W9s7u3ul/YOGjlPFsM5iEauHgGoUXGLdcCPwIVFIo0BgMxheT/zmIyrNY3lvRgl2ItqXPOSMGivdPX9bqnsVbwpyDLxc1KGHLVu6afdi1kaoTRMUK1bvpeYTkaV4Uzg2G2nGhPKhrSPLUsljVB3sumpY3JilR4JY2VLGjJV/05kNJ6FAW2M6JmoBe9ifif10pNeNnJuExSg5LNFoWpICYmk79JjytkRowsoUxeythA6oMzYd153bIznD0Fpjm42/mMQyaZxVfK/i356Xq1d5SkU4gmM4BR8uoAo3UIM6MOjDC7zCm/PsvDsfzuesteDkM4cwB+frFxdeluU=</latexit> <latexit sha1_base64="d96mskUPBfSUp9ih23g0ari3WA=">ACAXicbVA9SwNBEJ2LX/H8ilraLAbBKtyJoGXQxjKiSYQkhL3NXLJkb+/Y3VPCkcraVn+Dndj6S/wJ/gs3yRUm8cHA470ZuYFieDaeN63U1hZXVvfKG6W9s7u3ul/YOGjlPFsM5iEauHgGoUXGLdcCPwIVFIo0BgMxheT/zmIyrNY3lvRgl2ItqXPOSMGivdPXV5t1T2Kt4UZJn4OSlDjlq39NPuxSyNUBomqNYt30tMJ6PKcCZw7LZTjQlQ9rHlqWSRqg72fTUMTmxSo+EsbIlDZmqfycyGmk9igLbGVEz0IveRPzPa6UmvOxkXCapQclmi8JUEBOTyd+kxUyI0aWUKa4vZWwAVWUGZuO687tkZxhaK2xzcZfTGKZNM4qvlfxb8/L1as8pSIcwTGcg8XUIUbqEdGPThBV7hzXl23p0P53PWnDymUOYg/P1C3FGlx0=</latexit> <latexit sha1_base64="d96mskUPBfSUp9ih23g0ari3WA=">ACAXicbVA9SwNBEJ2LX/H8ilraLAbBKtyJoGXQxjKiSYQkhL3NXLJkb+/Y3VPCkcraVn+Dndj6S/wJ/gs3yRUm8cHA470ZuYFieDaeN63U1hZXVvfKG6W9s7u3ul/YOGjlPFsM5iEauHgGoUXGLdcCPwIVFIo0BgMxheT/zmIyrNY3lvRgl2ItqXPOSMGivdPXV5t1T2Kt4UZJn4OSlDjlq39NPuxSyNUBomqNYt30tMJ6PKcCZw7LZTjQlQ9rHlqWSRqg72fTUMTmxSo+EsbIlDZmqfycyGmk9igLbGVEz0IveRPzPa6UmvOxkXCapQclmi8JUEBOTyd+kxUyI0aWUKa4vZWwAVWUGZuO687tkZxhaK2xzcZfTGKZNM4qvlfxb8/L1as8pSIcwTGcg8XUIUbqEdGPThBV7hzXl23p0P53PWnDymUOYg/P1C3FGlx0=</latexit> <latexit sha1_base64="d96mskUPBfSUp9ih23g0ari3WA=">ACAXicbVA9SwNBEJ2LX/H8ilraLAbBKtyJoGXQxjKiSYQkhL3NXLJkb+/Y3VPCkcraVn+Dndj6S/wJ/gs3yRUm8cHA470ZuYFieDaeN63U1hZXVvfKG6W9s7u3ul/YOGjlPFsM5iEauHgGoUXGLdcCPwIVFIo0BgMxheT/zmIyrNY3lvRgl2ItqXPOSMGivdPXV5t1T2Kt4UZJn4OSlDjlq39NPuxSyNUBomqNYt30tMJ6PKcCZw7LZTjQlQ9rHlqWSRqg72fTUMTmxSo+EsbIlDZmqfycyGmk9igLbGVEz0IveRPzPa6UmvOxkXCapQclmi8JUEBOTyd+kxUyI0aWUKa4vZWwAVWUGZuO687tkZxhaK2xzcZfTGKZNM4qvlfxb8/L1as8pSIcwTGcg8XUIUbqEdGPThBV7hzXl23p0P53PWnDymUOYg/P1C3FGlx0=</latexit> <latexit sha1_base64="d96mskUPBfSUp9ih23g0ari3WA=">ACAXicbVA9SwNBEJ2LX/H8ilraLAbBKtyJoGXQxjKiSYQkhL3NXLJkb+/Y3VPCkcraVn+Dndj6S/wJ/gs3yRUm8cHA470ZuYFieDaeN63U1hZXVvfKG6W9s7u3ul/YOGjlPFsM5iEauHgGoUXGLdcCPwIVFIo0BgMxheT/zmIyrNY3lvRgl2ItqXPOSMGivdPXV5t1T2Kt4UZJn4OSlDjlq39NPuxSyNUBomqNYt30tMJ6PKcCZw7LZTjQlQ9rHlqWSRqg72fTUMTmxSo+EsbIlDZmqfycyGmk9igLbGVEz0IveRPzPa6UmvOxkXCapQclmi8JUEBOTyd+kxUyI0aWUKa4vZWwAVWUGZuO687tkZxhaK2xzcZfTGKZNM4qvlfxb8/L1as8pSIcwTGcg8XUIUbqEdGPThBV7hzXl23p0P53PWnDymUOYg/P1C3FGlx0=</latexit> CONVOLUTIONAL NET y x w 1 One’s hot label 0 0 1 0 w i ball car hat grumpy cat f ( x ; w ) → y Neural loss function X k f ( x i ; w ) � y i k 2 L ( w ) = min w i Non-convex?

  5. VISUALIZING LOSS FUNCTIONS: FILTER NORMALIZATION Step 1: Find w minimizer 30 million dimensions

  6. VISUALIZING LOSS FUNCTIONS: FILTER NORMALIZATION u v Step 2: Random w directions u, v

  7. VISUALIZING LOSS FUNCTIONS: FILTER NORMALIZATION u v Step 3: u i u i · k w i k Filter w k u i k normalization Li, Xu, Taylor, Studer, G. “Visualizing the loss landscape of neural nets.”

  8. VISUALIZING LOSS FUNCTIONS: FILTER NORMALIZATION u v Step 4 Plot Li, Xu, Taylor, Studer, G. “Visualizing the loss landscape of neural nets.”

  9. 56 LAYER “VGG” NET CIFAR-10

  10. 56 LAYER NEURAL NET CIFAR-10

Recommend


More recommend