545298 research outputs found
Sort by
Towards Resource-Efficient Trustworthy Distributed Learning
Get PDFFederated learning (FL) is a collaborative training framework, where multiple edge-devices (users) jointly train a global model, without sharing their private data. Instead, users share the local gradients with a central server after training with their on-device dataset. Several major challenges limit the deployment of conventional FL protocols in real-world large-scale networks. For example, mobile users may drop out from the protocol due to having limited communication and computation resources. Moreover, a single global model may not be optimal for users with heterogeneous data distributions. Lastly, user data can be retrieved even from the shared local gradients through gradient inversion attacks.
Clustered FL is a recent distributed learning mechanism to tackle data heterogeneity in FL, which allows training of multiple global models concurrently, each designated for a cluster of users, i.e., users with similar data distributions. To reduce communication latency in FL, a notable approach is over-the-air aggregation, where users share the same spectrum for sending gradients, enabling gradient aggregation directly over the wireless medium. Gradient sparsification is another widely adopted technique to alleviate communication cost, where each user sends only a small fraction of the entire gradient. For resource-efficient training of large models, a popular approach is submodel learning, which allows users to train a portion of the global model based on their available computing resources and bandwidth. With the prevalence of large foundation models in practical applications, parameter-efficient fine-tuning has recently gained attention, where users only train a few lightweight modules while keeping the pretrained model frozen, significantly reducing resource consumption.
The dissertation addresses several key challenges associated with trustworthy deployment of FL in resource-limited settings. Firstly, we propose an over-the-air aggregation mechanism for clustered FL, which uses the wireless medium to align the local gradients from the same cluster during aggregation, enabling the server to later correctly decode the gradient aggregate for each cluster. The underlying motivation is to tackle data heterogeneity and communication bottleneck simultaneously in large-scale FL settings. Next, we discover that user privacy can be breached when conventional secure aggregation (SA) protocols are naively applied to clustered FL or resource-aware FL protocols. This is attributed to the server's access to auxiliary information like cluster identity (in clustered FL), selected gradient coordinates (in gradient sparsification) or selected submodel indices (in submodel learning) of the user. To prevent associated privacy threats, we propose novel SA protocols, which ensure the information-theoretic privacy of both the auxiliary information and local gradients. Finally, we investigate the privacy risks associated with parameter-efficient fine-tuning, which has been largely underexplored. Accordingly, we identify a novel gradient inversion attack arising from a maliciously crafted pretrained model and fine-tuning modules, which demonstrates high-fidelity data recovery from the shared fine-tuning gradients
On the Risks of Generative AI: Addressing Safety Alignment in Vision-Language Models and Robust Deepfake Detection under Noisy Conditions
Get PDFGenerative AI refers to AI models capable of creating human-like content, such as text, images, or audio by learning patterns from large datasets. Examples include models like GPT and DALL·E, with applications in content creation, design, and more. Their usage has surged in recent years, becoming increasingly integrated into our daily lives. Therefore, it is the need of the hour to ensure a safe usage of such tools.
This thesis presents two independent studies. The first focuses on examining safety alignment in open-source vision-language models (VLMs) such as LLaVA-1.5 and LLaMA 3.2 Vision under early-exiting from the image encoder conditions. The second addresses the development of robust deepfake detection classifiers capable of performing reliably under network-induced distortions. Both of which are briefly described below:
Vision-language models (VLMs) have improved significantly in their capabilities, but their complex architecture makes their safety alignment challenging. In this work, we reveal an uneven distribution of harmful information across the intermediate layers of the image encoder and show that skipping a certain set of layers and exiting early can increase the chance of the VLM generating harmful responses. We call it as "Image enCoder Early-exiT'' based vulnerability (ICET). Our experiments across three VLMs: LLaVA-1.5, LLaVA-NeXT, and Llama 3.2, show that performing early exits from the image encoder significantly increases the likelihood of generating harmful outputs. To tackle this, we propose a simple yet effective modification of the Clipped-Proximal Policy Optimization (Clip-PPO) algorithm for performing layer-wise multi-modal RLHF for VLMs. We term this as Layer-Wise PPO (L-PPO). We evaluate our L-PPO algorithm across three multi-modal datasets and show that it consistently reduces the harmfulness caused by early exits.
The rapid advancement of deep generative models has made deepfakes increasingly accessible and widespread, highlighting the need for robust detection methods to mitigate their potential harm. While deepfake detection has been extensively studied, a critical gap yet remains to be explored: How well do existing detectors perform under real-world conditions involving common corruptions such as compression, or distortions arising from network transmission, media storage, or post-processing? We show that state-of-the-art deepfake detectors are surprisingly fragile in practical settings, with performance dropping significantly under wide range corruptions ranging from simple additive noise to more complex ones like motion blur, jitter and JPEG compression. Since the diversity of corruptions is difficult to capture during training, we propose a novel defense mechanism that models corruption at test-time as a mixture of multi-modal distributions. While we choose a Mixture of Gaussian (MoG) in this work, our framework is flexible and supports other distribution families. Given a corrupted deepfake image or video, we employ a Mixture Density Network (MDN) to estimate the means, variances, and mixing coefficients of the MoG that approximates the underlying noise. Based on the inferred mixture components, we augment the base detector with the corresponding Gaussian noise-specific pre-trained modules and dynamically ensemble them to produce robust predictions. Experimental results demonstrate that our method consistently outperforms baseline detectors across various deepfake datasets and model architectures
From Filaments to Clusters: The Evolving Role of the Cosmic Web in Galaxy Evolution Across Cosmic Time
No full textThis thesis focuses on the impact of Large Scale Structures on galaxies’ evolutionary paths. I used photometric redshifts of galaxies across the ∼ 2 deg2 COSMOS field to: (1) construct a density catalog for a subsample of galaxies at 0.4 < z < 5; (2) assess the contribution of internal (mass) and external (environmental) quenching mechanisms in shaping the star formation history; and (3) extract various components of the cosmic web, including fields, filaments, and clusters, and study galaxies’ distribution and evolution across these structures. I developed a pipeline that reconstructs the density field in redshift slices of constant comoving width using the weighted Kernel Density Estimation (wKDE) approach, and extracts components of the cosmic web by employing the Multi-Scale Morphology Filter technique, based on Hessian analysis of the density field. My results show that 14%, 32%, and 54% of galaxies reside in clusters, filaments, and fields, respectively. Out to z ∼ 1, the average SFR and sSFR decline in extremely dense environments, such as clusters, and at the high-mass end of the distribution, mostly due to the presence of massive quiescent galaxies. This decline is sharper for central and satellite galaxies compared to isolated galaxies. At 1 < z < 2, the environmental dependence diminishes, while stellar mass remains the dominant factor regulating star formation activity. Beyond z ∼ 2, the sample is dominated by star-forming galaxies, and we observe a reversal of the trends seen in the local Universe, where the average SFR increases in rich environments, such as clusters and filaments, stronger for centrals and satellites than isolated galaxies. Both environmental and mass quenching efficiencies increase with stellar mass at all redshifts, with mass quenching being the dominant factor in massive galaxies. At z > 2, negative values of environmental quenching efficiency suggest that the fraction of star-forming galaxies in dense environments exceeds that in less-dense regions, likely due to greater availability of cold gas, higher merger rates, and tidal effects that trigger star formation activity. Over time, however, environmental processes progressively suppress star formation in dense regions, shaping the emergence of the quiescent galaxy population in dense environments
Epitaxial Growth and Electrical Detection of the Néel Vector in Antiferromagnetic Insulators
No full textAntiferromagnetic (AFM) materials hold great promise for spintronic applications because they offer robustness against external perturbations, absence of stray fields, and terahertz spin dynamics. However, their vanishing net magnetization poses fundamental challenges for electrical manipulation and detection of the Néel vector. In this dissertation, we investigate (i) epitaxial growth and characterization of prototypical AFM insulator Cr2O3 (chromia) and α-Fe2O3 (hematite) thin films, and (ii) spin current generation and electrical detection of the Néel vector via spin Hall magnetoresistance (SMR), anomalous Hall effects (AHE), spin Seebeck effect (SSE), and spin pumping.In Cr2O3, we identify a magnetostriction-induced background that mimics SMR, enabling a cleaner separation of genuine spin-transport signals. Using proximity-induced anomalous Hall readout in Cr2O3(0001)/Pt, we demonstrate electrical detection of a single-domain Néel vector across the spin-flop transition. We further show that SSE provides a sensitive probe of the Néel temperature and reveals that oxygen deficiency during thin-film growth suppresses spin-current transport. In α-Fe2O3, we control the Morin transition through film thickness and detect the associated magnetic configurations using SMR. From these results, we develop an understanding of anisotropic field suppression of the Morin transition and extract the strength of the Dzyaloshinskii-Moriya (DM) interaction in hematite. Systematic spin pumping and inductive antiferromagnetic resonance experiments further elucidate the field dependence of AFM magnon spectrum, while hard-axis SSE measurements reveal hysteresis associated with Néel-vector reversal driven by the DM interaction. Also, we present preliminary results from α-Fe2O3/TaIrTe4 heterostructures suggesting that hematite imprints its magnetic order information onto the angular dependence of the SMR of α-Fe2O3/TaIrTe4 heterostructure. Our results provide new insight into spin-current generation, transport, and detection in AFM insulators and highlight pathways toward their integration in future spintronic devices
Shaping the Future of Generative AI: Insights into Interpretability, Alignment, and Compositionality of Diffusion Models
Get PDFGenerative models, including autoregressive models, GANs, VAEs, and diffusion models, have demonstrated remarkable capabilities in synthesizing high-quality content across diverse modalities. Among these, diffusion models have emerged as a particularly powerful framework due to their stability and generative fidelity.In this work, we present an information-theoretic perspective on diffusion modeling by establishing a precise connection between data distributions and optimal denoising via the pointwise Information–Minimum Mean Square Error (I-MMSE) relation. This connection refines traditional variational training objectives, enabling more accurate density estimation for both continuous and discrete data. Furthermore, by integrating contrastive learning, our method enhances the model’s ability to approximate complex data distributions. We term this framework Information-Theoretic Diffusion (ITD), which not only improves density modeling but also provides a principled interpretation of cross-modal alignment—modeling mutual information between modalities (e.g., text and image) rather than relying on attention maps or CLIP scores. In addition, ITD enables improved mutual information estimation directly within the training process.To fully leverage the information in the training data while minimizing computational cost, we introduce a plug-and-play lightweight path-mixer module, enabling highly efficient diffusion model training for under $2000. We also explore a conditional autoregressive diffusion framework that progressively injects information for flexibly compositionable generation. Overall, this work bridges information theory and diffusion modeling, advancing both the theoretical foundations and practical performance of generative AI
The Problem of Parasitic Agency and Anger as a Real-Time Remedy
Get PDFThe literature on epistemic injustice has generally concerned itself with the following tasks: Identifying a new species of epistemic injustice; identifying the harms (especially epistemic harms) of the newly identified epistemic injustice; and proposing a “future-oriented” remedy for the newly identified epistemic injustice. In this dissertation, I depart from this script in two critical ways. First, I draw attention to a genus, rather than individual species, of epistemic injustice: the genus of unjustly backfiring action. The phenomenon of unjustly backfiring action can be characterized as a marginalized agent’s exercising of agency that, contrary to her intentions, harms her while benefiting the dominant agent(s) or the oppressive system. This genus includes a variety of epistemic injustices recognized in the literature, including those involving coerced speech acts, distorted speech acts, and non-communicative acts involving distorted intentions. Examining this genus has the payoff of illuminating a kind of hegemonic agency that has not heretofore received attention: parasitic agency that is exercised by the dominant agents when they bring about unjustly backfiring action through overriding and co-opting the agency of marginalized agents. Parasitic agency is harmful because it victim-blames marginalized agents for the backfiring consequences and this can profoundly discourage the marginalized agents as epistemic agents. Second, with this kind of hegemonic agency in full view, I offer a real-time remedy, as opposed to a future-oriented remedy, to mitigating the harms of parasitic agency. Through protecting (and restoring) one’s competence-enabling epistemic goods—such as intellectual confidence, intellectual self-trust, appropriately-calibrated attention, and energy—a real-time remedy protects (and restores) one’s epistemic competences when marginalized agents face exercises of parasitic agency. I argue that anger can be a real-time remedy. In short, in this dissertation, I introduce and analyze the problem of parasitic agency through examining unjustly backfiring action and offer a real-time remedy for its harms. My dissertation can be divided into two halves. In the first-half, I illuminate the nature and oppressive function of parasitic agency. In the second-half, I establish that a real-time remedy is needed to mitigate the harms of parasitic agency and show that anger can be that remedy
Theory-Guided Systems Software for Predictable Autonomous Computing
No full textThis dissertation presents theory-guided systems software techniques for achieving predictable and reliable performance in autonomous systems built on middleware frameworks like the Robot Operating System (ROS). Modern autonomous systems face significant challenges in managing shared resources among multiple real-time tasks, particularly when heterogeneous hardware accelerators are involved and operating conditions are dynamic. This work addresses these challenges through two major contributions. First, we develop PAAM (Priority-driven Accelerator Access Management), a coordinated framework that conceptualizes accelerator access as a service, providing analytically-bounded latency guarantees for critical tasks while maintaining system efficiency. PAAM demonstrates up to 91% reduction in end-to-end latency for time-critical processing chains in autonomous driving scenarios. Second, we present LaME (Latency Management Executor), an adaptive scheduling framework that dynamically controls resource allocation with theoretical performance guarantees. LaME employs a two-level approach combining priority-based scheduling for real-time chains with fairness guarantees for best-effort workloads, achieving up to 400% improvement in worst-case response times under adverse conditions. Both frameworks are designed as drop-in replacements for existing ROS 2 components, enabling immediate deployment in real-world systems. Extensive evaluation on embedded platforms demonstrates that our solutions effectively bridge the gap between theoretical real-time guarantees and practical system implementation, enabling autonomous systems to maintain predictable behavior even under resource contention and environmental uncertainties
Effects of an Online Professional Development Activity on School Psychologists’ Consultation Self-Efficacy and Attitudes Toward Evidence-Based Practices to Reduce Racial Disparities in Exclusionary Discipline Practices
Get PDFExclusionary discipline practices, such as suspension and expulsion, are widely used in schools to manage student behavior but have been consistently linked to negative academic and behavioral outcomes, including lower achievement, increased dropout rates, and heightened antisocial behavior. A substantial body of research has documented enduring racial disparities in the application of these disciplinary measures, disproportionately affecting students of color and contributing to systemic inequities. Despite decades of evidence highlighting this disproportionality, particularly for African American students, schools continue to face challenges in addressing these disparities, and recent federal efforts may hinder progress by limiting schools’ ability to act on this issue.This study examined the impact of a brief professional development (PD) activity on school psychologists’ knowledge and self-efficacy related to consultation practices aimed at addressing racial disparities in exclusionary discipline. A paired samples t-test revealed statistically significant improvements in both content knowledge (Cohen’s d = 0.54) and self-efficacy (Cohen’s d = 0.56), suggesting the PD had a moderate and meaningful effect on participants' readiness to engage in equity-centered consultation.Participant perceptions of the intervention were generally favorable across multiple dimensions. Data from the adapted Usage Rating Profile–Intervention Revised (URP-IR) indicated positive attitudes regarding the acceptability (M = 4.44), feasibility (M = 4.28), system climate (M = 4.37), and system support (M = 4.96) for implementing discipline-related equity interventions. However, participants noted concerns about the adequacy of existing approaches, implementation complexity, and misalignment between intervention goals and current school practices. These findings highlight both the promise and the challenges of promoting equity through school psychological consultation.While the study's small sample size, lack of expert review in PD development, and use of adapted measures present limitations, results provide early evidence that even brief, targeted PD may enhance school psychologists’ preparedness to support systemic change. Future research should expand upon these findings with more robust methodologies to explore how professional development can translate into sustainable reductions in disciplinary disproportionality
ImgKnock: Knockoffs Feature Selection for Image Data via Latent Representation Learning
Get PDFFeature selection is a crucial topic in high-dimensional statistics, particularly in various applications ranging from physics, economics, health, to many others. A critical objective is properly controlling the false discovery rate (FDR) to limit the selection of irrelevant features. For pixelated image data, traditional feature selection techniques can be used but they often struggle with high dimensionality, computational inefficiency, and procedural rigidity. This work leverages the use of latent representation learning and the knockoffs filter to succinctly identify key features while controlling for the FDR.
We propose ImgKnock, a novel pipeline that integrates deep latent representation learning with the model-X knockoff framework. The pipeline employs four steps: i) generates a reduced set of latent representative features of the image data using self-supervised learning; ii) generates valid knockoffs of the latent features; iii) runs knockoff selection with FDR control; iv) interprets the selected latent features. Using the handwritten MNIST digits and CIFAR-10 image datasets, we are able to control the FDR and get AUC metrics of 0.911 and 0.779 respectively using deep knockoffs. We apply the novel procedure to fundus images to better understand the important features that are associated with glaucoma, a persistent eye disease. Modifying our method to use second-order representative group knockoffs, we are able to control the FDR and return AUC metrics of 0.7478 and 0.7356 in simulated runs with the latent representations of the fundus images. We also develop several tools that monitor the adherence of knockoff properties and analyze the latent representations. Future research should focus on extending interpretability and enhancing feature selection for multi-label classification tasks.
This dissertation is organized as follows. Chapter 1 introduces the knockoffs framework and its application in feature selection, particularly for image data. Chapter 2 delves into the use of knockoffs for multinomial and image data, highlighting both successes and limitations that motivate ImgKnock using latent features. Chapter 3 details the methodology and rationale behind ImgKnock, supported by simulated experiments demonstrating FDR control. Chapter 4 showcases the use of ImgKnock on glaucoma detection, with several modifications. Finally, Chapter 5 summarizes the work and discusses future research directions
Combinatorial Screening of Tricyclic Peptides for Biomedical Applications
No full textPeptide therapeutics hold great promise but are often limited by the conformational flexibility and functional constraints of linear scaffolds. To overcome these challenges, we developed a modular synthetic strategy for constructing tricyclic peptide libraries using orthogonal on-resin cyclization via copper-catalyzed azide–alkyne cycloaddition (CuAAC) and ring-closing metathesis (RCM). Optimization of stereochemistry and reaction conditions enabled efficient on-resin cyclization with >40% yield across diverse sequences. A dual-library approach, incorporating monocyclic analogs, facilitated robust sequence recovery via tandem mass spectrometry and enabled efficient hit validation. Screening against epitopes from intrinsically disordered proteins (IDPs), including MYC and APOBEC3A, revealed target-specific enrichment patterns, confirming the ability of tricyclic scaffolds to access novel binding space typically inaccessible to linear peptides.Focused screening against the MYC epitope yielded eleven initial hits, with NTC48 identified for further optimization. Guided by molecular dynamics simulations, we developed NTC81—an analog featuring a modified RCM linkage that enhances both backbone rigidity and binding affinity (Kd ≈ 30 nM). Biophysical analyses, including ion mobility mass spectrometry and MD clustering, showed that NTC81 adopts a compact, preorganized conformation ideally suited for engaging the disordered HLH–LZ region of MYC. In cellular assays, NTC81 demonstrated membrane permeability, target engagement with endogenous MYC (via CETSA), disruption of MYC-driven transcriptional programs and spheroid morphology, and selective cytotoxicity in MYC-amplified cell lines. These findings establish tricyclic peptides as a powerful platform for targeting disordered protein interfaces and position NTC81 as a promising first-in-class MYC inhibitor.To further expand structural diversity and explore untapped regions of conformational space, we designed an asymmetric tricyclic peptide library. By combining orthogonal cyclization strategies with rational incorporation of biologically inspired motifs, we generated topologically complex scaffolds with increased 3D diversity and functional modularity. Despite synthetic challenges, this approach enabled the construction of multifunctional tricyclic peptides with broad potential—as both high-affinity binders to undruggable targets like MYC and as versatile tools for probing protein–protein interactions in chemical biology