Pre-trained Large Language Models (LLMs) are an integral part of modern AI that have led to breakthrough performances in complex tasks. Major companies with expensive infrastructures able develop and train these large models billions millions parameters from scratch. Third parties, researchers, practitioners increasingly adopting pre-trained fine-tuning them on their private data accomplish downstream However, it has been shown adversary can extract/reconstruct the exact training samples...

Federated learning (FL) enhances privacy by keeping user data on local devices. However, emerging attacks have demonstrated that the updates shared users during training can reveal significant information about their data. This has greatly thwart adoption of FL methods for robust AI models in sensitive applications. Differential Privacy (DP) is considered gold standard safeguarding DP guarantees are highly conservative, providing worst-case guarantees. result overestimating needs, which may...

Recent machine learning- and deep learning-based static malware detectors have shown breakthrough performance in identifying unseen variants. As a result, they are increasingly being adopted to lower the cost of dynamic analysis manual signature identification. Despite their success, studies that can be vulnerable adversarial attacks, which an adversary modifies known executable subtly fool detector into recognizing it as benign file. automatically crafting these variants at scale is...

Driving safety analysis has recently experienced unprecedented improvements thanks to technological advances in precise positioning sensors, artificial intelligence (AI)-based features, autonomous driving systems, connected vehicles, high-throughput computing, and edge computing servers. Particularly, deep learning (DL) methods empowered volume video processing extract safety-related features from massive videos captured by roadside units (RSU). Safety metrics are commonly used measures...

Articulated motion analysis often utilizes strong prior knowledge such as a known or trained parts model for humans. Yet, the world contains variety of articulating objects--mammals, insects, mechanized structures--where number and configuration particular object is unknown in advance. Here, we relax assumptions via an unsupervised, Bayesian nonparametric that infers with motions coupled by body dynamic parameterized SE(D), Lie group rigid transformations. We derive inference procedure short...

Empowered by the recent development in Ma-chine Learning (ML), signatureless ML-based malware detectors present promising performance identifying unseen mal ware variants and zero days without requiring expensive dynamic analysis. However, it has been recently shown that are vulnerable to adversarial attacks, which an attacker modifies a known exe-cutable trick detector into recognizing modi-fied variant as benign. Adversarial example generation become emerging area ML studies creating...

Artificial Intelligence (AI) is being widely adopted in modern cyber defense to weave automation and scalability into the operational fabric of cybersecurity firms. Today, AI aids crucial tasks such as malware intrusion detection keep Information Technology (IT) infrastructure secure. Despite their value, agents can be vulnerable adversarial attacks. In these attacks, adversary deliberately manipulates a malicious input by taking sequence actions so that targeted agent fails correctly...

Differentially private stochastic gradient descent (DP-SGD) has been instrumental in privately training deep learning models by providing a framework to control and track the privacy loss incurred during training. At core of this computation lies subsampling method that uses amplification lemma enhance guarantees provided additive noise. Fixed size is appealing for its constant memory usage, unlike variable sized minibatches Poisson subsampling. It also interest addressing class imbalance...

Driving safety analysis has recently witnessed unprecedented results due to advances in computation frameworks, connected vehicle technology, new generation sensors, and artificial intelligence (AI). Particularly, the recent performance of deep learning (DL) methods realized higher levels for autonomous vehicles empowered volume imagery processing driving analysis. An important application DL is extracting metrics from traffic imagery. However, majority current use micro-scale individual...

We apply the expectation propagation (EP) algorithm to temporally track targets using sensors that produce spurious clutter detections, and may sometimes fail detect true target. The variational inference framework underlying EP allows tracker be easily adapted varying measurement models. develop variants of based on single Gaussian mixture approximations posterior target location distributions, which offer a tradeoff between accuracy computational complexity. Experiments show improved...

for the U. S. Nuclear Regulatory Commission’s Advisory Committee on Reactor Safeguards in 2003. That report noted that “in-dustry lacks tools to perform time-trend analysis with Bayesian updating.” This paper describes an applica-tion of time-dependent inference methods developed European Commission Ageing PSA Network. These utilize open-source software, implementing Markov chain Monte Carlo sampling. The also illustrates approach incorporating multiple sources data via applicability...

We present an approach to data fusion that combines the interpretability of structured probabilistic graphical models with flexibility neural networks. The proposed method, lightweight (LDF), emphasizes posterior analysis over latent variables using two types information: primary data, which are well-characterized but limited availability, and auxiliary readily available lacking a statistical relationship quantity interest. lack forward model for precludes use standard approaches, while...

Pre-trained Large Language Models (LLMs) are an integral part of modern AI that have led to breakthrough performances in complex tasks. Major companies with expensive infrastructures able develop and train these large models billions millions parameters from scratch. Third parties, researchers, practitioners increasingly adopting pre-trained fine-tuning them on their private data accomplish downstream However, it has been shown adversary can extract/reconstruct the exact training samples...