Hindering Adversarial Attacks with Implicit Neural Representations

Authors: Andrei A Rusu, Dan Andrei Calian, Sven Gowal, Raia Hadsell

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We introduce the Lossy Implicit Network Activation Coding (LINAC) defence, an input transformation which successfully hinders several common adversarial attacks on CIFAR-10 classifiers for perturbations up to ϵ = 8/255 in L∞ norm and ϵ = 0.5 in L2 norm.
Researcher Affiliation Industry 1Deep Mind, London, UK. Correspondence to: Andrei A. Rusu <andrei@deepmind.com>.
Pseudocode Yes We provide pseudocode for the LINAC transform in Algorithm 1 and a discussion of computational and memory requirements in Appendix A.1.4.
Open Source Code No (*) Source code will be released before acceptance.
Open Datasets Yes We introduce the Lossy Implicit Network Activation Coding (LINAC) defence, an input transformation which successfully hinders several common adversarial attacks on CIFAR-10 classifiers
Dataset Splits No The paper mentions using CIFAR-10 training and test sets but does not explicitly provide specific numerical percentages or sample counts for training, validation, and test splits, nor does it cite a predefined split.
Hardware Specification No The paper mentions 'modern SIMD devices' and references to computational cost comparable to 'Wide Res Net-70-16' inference, but does not provide specific hardware details such as GPU or CPU models, or memory specifications used for the experiments.
Software Dependencies No The paper mentions software libraries like JAX and NumPy by citing their respective papers, but it does not provide specific version numbers for these dependencies.
Experiment Setup Yes Fitting the parameters of the implicit neural network was done using Adam (Kingma & Ba, 2015), with default parameters and a learning rate µ = 0.001. We used mini-batches with M = 32 random pixels and trained for N = 10 epochs. An epoch constitutes a pass through the entire set of pixels in the input image with dimensions I J C = 32 32 3 in random order. The total number of optimisation steps performed was 320. A cosine learning rate decay schedule was used for better convergence, with the minimum value of the multiplier α = 0.0001 (Loshchilov & Hutter, 2016). (...) Training was performed with Nesterov Momentum SGD (Tieleman & Hinton, 2012) m = 0.9, using mini-batches of size 1024, for a total of 1000 epochs, or 48880 parameter updates. The initial learning rate was µ = 0.4, reduced by a factor of 10 four times, at epochs: 650, 800, 900 and 950. We performed a hyper-parameter sweep over the weight-decay scale with the following grid: {0., 0.0001, 0.0005, 0.0010}.