Compression techniques for deep learning have become increasingly popular, particularly in settings where latency and memory constraints are imposed.Several methods, such as pruning, distillation, and quantization, have been adopted for compressing models, each providing distinct advantages.However, existing literature demonstrates that compressing deep learning models could affect their fairness.Our analysis involves a comprehensive evaluation of pruned, distilled, and quantized language models, which we benchmark across a range of intrinsic and extrinsic metrics for measuring bias in text classification.We also investigate the impact of using multilingual models and evaluation measures.Our findings highlight the significance of considering both the pre-trained model and the chosen compression strategy in developing equitable language technologies.The results also indicate that compression strategies can have an adverse effect on fairness measures.

Kabir Ahuja, Harshita Diddee, Rishav Hada, Millicent Ochieng, Krithika Ramesh, Prachi Jain, Akshay Nambi, Tanuja Ganu, Sameer Segal, Mohamed Ahmed, Kalika Bali, Sunayana Sitaram. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023.

Privacy-preserving machine learning (PPML) promises to train machine learning (ML) models by combining data spread across multiple data silos. Theoretically, secure multiparty computation (MPC) allows multiple data owners to train models on their joint data without revealing the data to each other. However, the prior implementations of this secure training using MPC have three limitations: they have only been evaluated on CNNs, and LSTMs have been ignored; fixed point approximations have affected training accuracies compared to training in floating point; and due to significant latency overheads of secure training via MPC, its relevance for practical tasks with streaming data remains unclear. The motivation of this work is to report our experience of addressing the practical problem of secure training and inference of models for urban sensing problems, e.g., traffic congestion estimation, or air pollution monitoring in large cities, where data can be contributed by rival fleet companies while balancing the privacy-accuracy trade-offs using MPC-based techniques.Our first contribution is to design a custom ML model for this task that can be efficiently trained with MPC within a desirable latency. In particular, we design a GCN-LSTM and securely train it on time-series sensor data for accurate forecasting, within 7 minutes per epoch. As our second contribution, we build an end-to-end system of private training and inference that provably matches the training accuracy of cleartext ML training. This work is the first to securely train a model with LSTM cells. Third, this trained model is kept secret-shared between the fleet companies and allows clients to make sensitive queries to this model while carefully handling potentially invalid queries. Our custom protocols allow clients to query predictions from privately trained models in milliseconds, all the while maintaining accuracy and cryptographic security.

Over the last few years, YouTube Kids has emerged as one of the highly competitive alternatives to television for children's entertainment. Consequently, YouTube Kids' content should receive an additional level of scrutiny to ensure children's safety. While research on detecting offensive or inappropriate content for kids is gaining momentum, little or no current work exists that investigates to what extent AI applications can (accidentally) introduce content that is inappropriate for kids. In this paper, we present a novel (and troubling) finding that well-known automatic speech recognition (ASR) systems may produce text content highly inappropriate for kids while transcribing YouTube Kids' videos. We dub this phenomenon as inappropriate content hallucination. Our analyses suggest that such hallucinations are far from occasional, and the ASR systems often produce them with high confidence. We release a first-of-its-kind data set of audios for which the existing state-of-the-art ASR systems hallucinate inappropriate content for kids. In addition, we demonstrate that some of these errors can be fixed using language models.

Social media sites such as Twitter have given a platform to people to express their thoughts even on the most sensitive of topics. In this paper, we analyze tweets which are posted by people who acknowledged their sexuality and shared their friends and families' response publicly. In our experiment, we have used a set of Tweets data with specific hashtags to analyze the response. Context-based topic modeling was used to gather tweets surrounding this topic to get the relevant tweets. The unwanted tweets discovered by the topic model was discarded. The sentiment of these tweets was then extracted using novel methods. The motivation for this process was to find out how their declaration of sexuality was perceived. Unfortunately, homophobia is one of the battles humanity continues to fight.