I earned my Ph.D. in computer science from Florida International University (FIU) in December 2023, advised by Dr. Niki Pissinou. I also earned an M.S. in telecommunications and networking from FIU in December 2022. My graduate studies were supported by the GEM Consortium with MIT Lincoln Laboratory, an FIU Director's Fellowship, and a 2021 NSF Graduate Research Fellowship.
My dissertation, Mitigating Adversarial Perturbations in Machine Learning Models in the Presence of Uncertainty, specialized in adversarial machine learning, with the goal of defending against imperceptible adversarial perturbations that exploit model uncertainty. I analyzed learning system components (training data transformations, activation function characteristics, and overall model complexity) to modify the loss surface and increase adversarial robustness. From 2019, I also co-advised projects in adversarial machine learning for domain-specific applications such as the Internet of Things (IoT) and natural language processing (NLP) through the NSF-funded Research Experience for Undergraduates and Research Experience for Teachers programs.
Increasing adversarial robustness around uncertain boundary regions with amodal segmentation
IEEE International Conference on Machine Learning and Applications proceedings (ICMLA 2024)
abstract
Adversarial perturbations in object recognition and image classification tasks impact a learning model's ability to perform accurately and increase the safety risks of deployed machine learning models. During the adversarial example generation process, adversaries approach areas most prone to model uncertainty. Identifying partially occluded items, especially without understanding general object shapes, contributes to significant model uncertainty since object boundaries are not inherently at the forefront of the feature generalization process in deep learning models. Thus, this work aims to reduce model uncertainty surrounding partially occluded boundaries and increase adversarial robustness by augmenting the training dataset with amodal segmentation boundary masks. By observing performance degradation, robust sensitivity, and loss sensitivity, we show how including these masks during training impacts an adversary's ability to generate effective adversarial examples on the versatile MS COCO dataset. Lastly, we observe how including these masks during training influences the performance of adversarial training.
Unifying robust activation functions for reduced adversarial vulnerability with the parametric generalized gamma function
IEEE International Conference on Machine Learning and Applications proceedings (ICMLA 2024)
abstract code
Adversaries minimally perturb deep learning input data to reduce a learning model's ability to produce domain-specific data-driven recommendations to solve specialized tasks. This vulnerability to adversarial perturbations has been argued to stem from a learning model's nonlocal generalization over complex input data. Given the incomplete information in a complex dataset, a learning model captures nonlinear patterns between data points with volatility in the loss surface and exploitable areas of low-confidence knowledge. It is the responsibility of activation functions to capture the nonlinearity in data and, thus, has inspired disjointed research efforts to create robust activation functions. This work unifies the properties of activation functions that contribute to robust generalization with the generalized gamma distribution function. We show that combining the disjointed characteristics presented in the literature with our parametric generalized gamma activation function provides more effective robustness than the individual characteristics alone.
Towards parametric robust activation functions in adversarial machine learning
Tiny Paper at International Conference on Learning Representations (ICLR 2023)
abstract code
Machine learning’s vulnerability to adversarial perturbations has been argued to stem from a learning model’s non-local generalization over complex input data. Given the incomplete information in a complex dataset, a learning model captures non-linear patterns between data points with volatility in the loss surface and exploitable areas of low-confidence knowledge. It is the responsibility of activation functions to capture the non-linearity in data and, thus, has inspired disjointed research efforts to create robust activation functions. This work unifies the properties of activation functions that contribute to robust generalization with the generalized gamma distribution function. We show that combining the disjointed characteristics presented in the literature provides more effective robustness than the individual characteristics alone. (Extended paper: Unifying robust activation functions for reduced adversarial vulnerability with the parametric generalized gamma function)
The dilemma between data transformations and adversarial robustness in time series application systems
AAAI's Workshop on Artificial Intelligence Safety proceedings (Safe AI 2022)
abstract
Adversarial examples, or nearly indistinguishable inputs created by an attacker, significantly reduce machine learning accuracy. Theoretical evidence has shown that the high intrinsic dimensionality of datasets facilitates an adversary's ability to develop effective adversarial examples in classification models. Adjacently, the presentation of data to a learning model impacts its performance. For example, we have seen this through dimensionality reduction techniques used to aid with the generalization of features in machine learning applications. Thus, data transformation techniques go hand-in-hand with state-of-the-art learning models in decision-making applications such as intelligent medical or military systems. With this work, we explore how data transformations techniques such as feature selection, dimensionality reduction, or trend extraction techniques may impact an adversary's ability to create effective adversarial samples on a recurrent neural network. Specifically, we analyze it from the perspective of the data manifold and the presentation of its intrinsic features. Our evaluation empirically shows that feature selection and trend extraction techniques may increase the RNN's vulnerability. A data transformation technique reduces the vulnerability to adversarial examples only if it approximates the dataset's intrinsic dimension, minimizes codimension, and maintains higher manifold coverage.
Jespipe: A plugin-based, open MPI framework for adversarial machine learning analysis
IEEE International Conference on Big Data proceedings (BigData 2021)
abstract code
Research is increasingly showing the tremendous vulnerability in machine learning models to seemingly undetectable adversarial inputs. One of the current limitations in adversarial machine learning research is the incredibly time-consuming testing of novel defenses against various attacks and across multiple datasets, even with high computing power. To address this limitation, we have developed Jespipe as a new plugin-based, parallel-by-design Open MPI framework that aids in evaluating the robustness of machine learning models. The plugin-based nature of this framework enables researchers to specify any pre-training data manipulations, machine learning models, adversarial models, and analysis or visualization metrics with their input Python files. Because this framework is plugin-based, a researcher can easily incorporate model implementations using popular deep learning libraries such as PyTorch, Keras, TensorFlow, Theano, or MXNet, or adversarial robustness tools such as IBM's Adversarial Robustness Toolbox or Foolbox. The parallelized nature of this framework also enables researchers to evaluate various learning or attack models with multiple datasets simultaneously by specifying all the models and datasets they would like to test with our XML control file template. Overall, Jespipe shows promising results by reducing latency in adversarial machine learning algorithm development and testing compared to traditional Jupyter notebook workflows.