An official website of the United States government
In recent news, organizations have been considering the use of facial and emotion recognition for applications involving youth such as tackling surveillance and security in schools. However, the majority of efforts on facial emotion recognition research have focused on adults. Children, particularly in their early years, have been shown to express emotions quite differently than adults. Thus, before such algorithms are deployed in environments that impact the wellbeing and circumstance of youth, a careful examination should be made on their accuracy with respect to appropriateness for this target demographic. In this work, we utilize several datasets that contain facial expressions of children linked to their emotional state to evaluate eight different commercial emotion classification systems. We compare the ground truth labels provided by the respective datasets to the labels given with the highest confidence by the classification systems and assess the results in terms of matching score (TPR), positive predictive value, and failure to compute rate. Overall results show that the emotion recognition systems displayed subpar performance on the datasets of children's expressions compared to prior work with adult datasets and initial human ratings. We then identify limitations associated with automated recognition of emotions in children and provide suggestions on directions with enhancing recognition accuracy through data diversification, dataset accountability, and algorithmic regulation.
more »
« less
This content will become publicly available on October 1, 2025
Zero-Shot Demographically Unbiased Image Generation From an Existing Biased StyleGAN
Face recognition systems have made significant strides thanks to data-heavy deep learning models, but these models rely on large privacy-sensitive datasets. Recent work in facial analysis and recognition have thus started making use of synthetic datasets generated from GANs and diffusion based generative models. These models, however, lack fairness in terms of demographic representation and can introduce the same biases in the trained downstream tasks. This can have serious societal and security implications. To address this issue, we propose a methodology that generates unbiased data from a biased generative model using an evolutionary algorithm. We show results for StyleGAN2 model trained on the Flicker Faces High Quality dataset to generate data for singular and combinations of demographic attributes such as Black and Woman. We generate a large racially balanced dataset of 13.5 million images, and show that it boosts the performance of facial recognition and analysis systems whilst reducing their biases. We have made our code-base ( https://github.com/anubhav1997/youneednodataset ) public to allow researchers to reproduce our work.
more »
« less
- Award ID(s):
- 1956200
- PAR ID:
- 10559763
- Publisher / Repository:
- IEEE
- Date Published:
- Journal Name:
- IEEE Transactions on Biometrics, Behavior, and Identity Science
- Volume:
- 6
- Issue:
- 4
- ISSN:
- 2637-6407
- Page Range / eLocation ID:
- 498 to 514
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
While machine learning approaches to visual emotion recognition o er great promise, current methods consider training and testing models on small scale datasets covering limited visual emotion concepts. Our analysis identi es an important but long overlooked issue of existing visual emotion benchmarks in the form of dataset biases. We design a series of tests to show and measure how such dataset biases obstruct learning a generalizable emotion recognition model. Based on our analysis, we propose a webly supervised approach by leveraging a large quantity of stock image data. Our approach uses a simple yet e ective curriculum guided training strategy for learning discriminative emotion features. We discover that the models learned using our large scale stock image dataset exhibit signi cantly better generalization ability than the existing datasets without the manual collection of even a single label. Moreover, visual representation learned using our approach holds a lot of promise across a variety of tasks on di erent image and video datasets.more » « less
-
Agents must monitor their partners' affective states continuously in order to understand and engage in social interactions. However, methods for evaluating affect recognition do not account for changes in classification performance that may occur during occlusions or transitions between affective states. This paper addresses temporal patterns in affect classification performance in the context of an infant-robot interaction, where infants’ affective states contribute to their ability to participate in a therapeutic leg movement activity. To support robustness to facial occlusions in video recordings, we trained infant affect recognition classifiers using both facial and body features. Next, we conducted an in-depth analysis of our best-performing models to evaluate how performance changed over time as the models encountered missing data and changing infant affect. During time windows when features were extracted with high confidence, a unimodal model trained on facial features achieved the same optimal performance as multimodal models trained on both facial and body features. However, multimodal models outperformed unimodal models when evaluated on the entire dataset. Additionally, model performance was weakest when predicting an affective state transition and improved after multiple predictions of the same affective state. These findings emphasize the benefits of incorporating body features in continuous affect recognition for infants. Our work highlights the importance of evaluating variability in model performance both over time and in the presence of missing data when applying affect recognition to social interactions.more » « less
-
null (Ed.)Abstract People spontaneously infer other people’s psychology from faces, encompassing inferences of their affective states, cognitive states, and stable traits such as personality. These judgments are known to be often invalid, but nonetheless bias many social decisions. Their importance and ubiquity have made them popular targets for automated prediction using deep convolutional neural networks (DCNNs). Here, we investigated the applicability of this approach: how well does it generalize, and what biases does it introduce? We compared three distinct sets of features (from a face identification DCNN, an object recognition DCNN, and using facial geometry), and tested their prediction across multiple out-of-sample datasets. Across judgments and datasets, features from both pre-trained DCNNs provided better predictions than did facial geometry. However, predictions using object recognition DCNN features were not robust to superficial cues (e.g., color and hair style). Importantly, predictions using face identification DCNN features were not specific: models trained to predict one social judgment (e.g., trustworthiness) also significantly predicted other social judgments (e.g., femininity and criminal), and at an even higher accuracy in some cases than predicting the judgment of interest (e.g., trustworthiness). Models trained to predict affective states (e.g., happy) also significantly predicted judgments of stable traits (e.g., sociable), and vice versa. Our analysis pipeline not only provides a flexible and efficient framework for predicting affective and social judgments from faces but also highlights the dangers of such automated predictions: correlated but unintended judgments can drive the predictions of the intended judgments.more » « less
-
Chatbot systems have improved significantly because of the advances made in language modeling. These machine learning systems follow an end-to-end data-driven learning paradigm and are trained on large conversational datasets. Imperfections or harmful biases in the training datasets can cause the models to learn toxic behavior, and thereby expose their users to harmful responses. Prior work has focused on measuring the inherent toxicity of such chatbots, by devising queries that are more likely to produce toxic responses. In this work, we ask the question: How easy or hard is it to inject toxicity into a chatbot after deployment? We study this in a practical scenario known as Dialog-based Learning (DBL), where a chatbot is periodically trained on recent conversations with its users after deployment. A DBL setting can be exploited to poison the training dataset for each training cycle. Our attacks would allow an adversary to manipulate the degree of toxicity in a model and also enable control over what type of queries can trigger a toxic response. Our fully automated attacks only require LLM-based software agents masquerading as (malicious) users to inject high levels of toxicity. We systematically explore the vulnerability of popular chatbot pipelines to this threat. Lastly, we show that several existing toxicity mitigation strategies (designed for chatbots) can be significantly weakened by adaptive attackers.more » « less
