Attribution-Based Confidence Metric For Deep Neural Networks

Authors: Susmit Jha, Sunny Raj, Steven Fernandes, Sumit K. Jha, Somesh Jha, Brian Jalaian, Gunjan Verma, Ananthram Swami

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We mathematically motivate the proposed metric and evaluate its effectiveness with two sets of experiments. First, we study the change in accuracy and the associated confidence over out-of-distribution inputs. Second, we consider several digital and physically realizable attacks such as FGSM, CW, Deep Fool, PGD, and adversarial patch generation methods. The ABC metric is low on out-of-distribution data and adversarial examples, where the accuracy of the model is also low. These experiments demonstrate the effectiveness of the ABC metric towards creating more trustworthy and resilient DNNs.
Researcher Affiliation Collaboration Susmit Jha Computer Science Laboratory SRI International Sunny Raj, Steven Lawrence Fernandes, Sumit Kumar Jha Computer Science Department University of Central Florida, Orlando Somesh Jha University of Wisconsin-Madison and Xaipient Brian Jalaian, Gunjan Verma, Ananthram Swami US Army Research Laboratory Adelphi
Pseudocode Yes Algorithm 1 Evaluate ABC confidence metric c(F, x) of machine learning model F on input x Input: Model F, Input x with features x1, x2, . . . xn, Sample size S Output: ABC metric c(F, x) 1: A1, . . . An Attributions of features x1, x2, . . . xn from input x 2: i F(x) {Obtain model prediction} 3: for j = 1 to n do 4: P(xj) |Aj/xj| Pn k=1 |Ak/xk| 5: end for 6: Generate S samples by mutating feature xj of input x to baseline xb j with probability P(xj) 7: Obtain the output of the model on the S samples. 8: c(F, x) Sconform/S where model s output on Sconform samples is i 9: return c(F, x) as confidence metric (ABC) of prediction by the model F on the input x
Open Source Code No The paper does not provide an explicit statement or link to the open-source code for the proposed ABC metric or any other methodology described.
Open Datasets Yes We empirically evaluate the ABC metric over MNIST and Image Net datasets using (a) out-of-distribution data, (b) adversarial inputs generated using digital attacks such as FGSM, PGD, CW and Deep Fool, and (c) physically-realizable adversarial patches and La VAN attacks. Out-of-distribution data: MNIST [65] with rotation and background, not MNIST [66] and Fashion MNIST [67].
Dataset Splits No The paper mentions 'held-out validation set' in the context of related work on calibration models but does not provide specific train/validation/test dataset splits for its own experiments.
Hardware Specification Yes All experiments were conducted on a 8 core Intel Core i9-9900K 3.60GHz CPU with NVIDIA Titan RTX Graphics and 32 GB RAM.
Software Dependencies No The paper mentions using 'Cleverhans' but does not specify any version numbers for this or other software dependencies.
Experiment Setup No The paper does not provide specific experimental setup details such as hyperparameter values (e.g., learning rate, batch size, number of epochs), optimizer settings, or other system-level training configurations.