Judicial review and remedies for algorithmic bias in the application of environmental law

Abstract

In this paper, the definitions of bias and fairness in deep learning are systematically sorted out, and the specificity of algorithmic bias in environmental law application scenarios is clarified. An algorithmic fairness testing method EIDIG based on gradient search is proposed for uncovering and correcting algorithmic bias.EIDIG adopts the gradient of the model output to replace the gradient of the loss function, which reduces the computational load of the model. Combined with the clustering algorithm to generate diverse individual bias samples as inputs for the next stage, the global search is completed. In the local search, the generated individual bias samples are imported and repeated detection is performed around them. The experimental results show that compared with the comparative fairness testing methods such as ADF and AEQUITAS, the number of bias samples generated by EIDIG, the success rate, and the generation efficiency are all improved to different degrees, and the time taken by EIDIG to generate 1,000 bias samples is only 43.90% and 88.06% of that taken by AEQUITAS and ADF, and EIDIG has achieved a leading position in improving the original model's fairness by achieving leading performance. Finally, this paper proposes judicial remedy strategies such as adjusting the civil law tort liability framework and introducing the public interest litigation system. It provides theoretical support for realizing the rule of law guarantee for environmental disputes in the era of artificial intelligence.