Busca avançada
Ano de início
Entree


Investigating Universal Adversarial Attacks Against Transformers-Based Automatic Essay Scoring Systems

Texto completo
Autor(es):
Silveira, Igor Cataneo ; Barbosa, Andre ; da Costa, Daniel Silva Lopes ; Maua, Denis Deratani
Número total de Autores: 4
Tipo de documento: Artigo Científico
Fonte: INTELLIGENT SYSTEMS, BRACIS 2024, PT II; v. 15413, p. 15-pg., 2025-01-01.
Resumo

Automatic Essay Scoring promises to scale up student feedback on written input, addressing the excessive cost and time demand associated with human grading. State-of-the-art automatic scorers are based on Transformers-based neural networks. While such models have shown impressive results in reasoning tasks, learned models often produce answers that arise from statistical clues in datasets and are misaligned with human objectives. Such systems are thus potentially fragile for scenarios where users are incentivized to deceive the system, as in a classroom setting. In this work, we evaluate the susceptibility of state-of-the-art automatic scorers to attacks made by non-expert users, such as students interacting with an automatic grader. We develop a methodology to simulate such student attacks and test them against scorers based on BERT, Phi-3 and Gemini models. Our findings suggest that (i) a BERT-based grader can be deceived using simple feature-based attacks; (ii) although Google's Gemini has a solid agreement with graders, it can assign undeservedly high grades for small sentences; (iii) a Phi-3-based grader was less susceptible than BERT, but it still assigned relatively high grades to some of our attacks. (AU)

Processo FAPESP: 19/07665-4 - Centro de Inteligência Artificial
Beneficiário:Fabio Gagliardi Cozman
Modalidade de apoio: Auxílio à Pesquisa - Programa eScience e Data Science - Centros de Pesquisa em Engenharia
Processo FAPESP: 22/02937-9 - Indução de programas lógico-probabilístico-neurais
Beneficiário:Denis Deratani Mauá
Modalidade de apoio: Auxílio à Pesquisa - Projeto Inicial