A repo for the Causal Subobject Explanations algorithm.
Automatic obfuscation of regions in images using machine learning explanation has emerged as a crit- ical problem in recent times. However, explanation for image obfuscation demands pinpointing spe- cific subobject regions of the input that are causal of a model’s particular decision, because subobject regions are human-understandable features, and causal explanation enables a user to understand what region caused a prediction for which it was masked. Existing explanation approaches can only provide attributions at a feature level and cannot address the need for region-level causal explanations. In this work, we propose a technique called Causal Subobject Explanations (CSE) that are counterfactual explanations produced based on an adaptive region binary masking algorithm with region attribution score heuristics, to identify the regions in the image that caused the model’s prediction to change. Extensive experiments on a baseline dataset demonstrates the effectiveness of CSE compared with five state-of-the-art ex- plainability and four clustering approaches in terms of three evaluation metrics. Furthermore, we demonstrate the practicality of CSEs for three datasets of harmful images, to automatically obfuscate the harmful regions in these images, thereby rendering them safe.
Requires a build from https://github.com/BGU-CS-VIL/BASS. This build is already provided if you clone this repo (bass_build).