BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//wp-events-plugin.com//7.4.8//EN
TZID:Asia/Kolkata
X-WR-TIMEZONE:Asia/Kolkata
BEGIN:VEVENT
UID:229@cds.iisc.ac.in
DTSTART;TZID=Asia/Kolkata:20261023T110000
DTEND;TZID=Asia/Kolkata:20261023T120000
DTSTAMP:20261016T144456Z
URL:https://cds.iisc.ac.in/events/ph-d-thesis-colloquium-102-cds-23-octobe
 r-2026-collaboration-amid-differences-generalization-adaptation-incentiviz
 ed-participation-and-minimax-optimization-in-federated-learning/
SUMMARY:Ph.D: Thesis Colloquium: 102: CDS: 23\, October 2026 "Collaboration
  amid Differences: Generalization\, Adaptation\, Incentivized Participatio
 n and Minimax Optimization in Federated Learning"
DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nPh.D. Thesis Col
 loquium\n\n\n\nSpeaker: Mr. Yashwanth. M\nS.R. Number: 06-18-01-10-12-21-1
 -19272\nTitle: "Collaboration amid Differences: Generalization\, Adaptatio
 n\, Incentivized Participation and Minimax Optimization in Federated Learn
 ing"\nResearch Supervisor: Prof. Anirban Chakraborty\nDate &amp\; Time : O
 ctober 23\, 2026 (Friday)\, 11:00 AM\nVenue : #102\, CDS Seminar Hall\n\n\
 n\nABSTRACT\n\nFederated Learning (FL) is a distributed machine learning p
 aradigm\, where multiple clients train a model collaboratively under the o
 rchestration of a central server while keeping the training data private t
 o respective clients. Since its introduction\, FL has\nbeen enabling numer
 ous privacy-preserving applications across major industries like healthcar
 e and big-tech. Despite its advantages\, the federated framework introduce
 s several critical algorithmic challenges under practical constraints. Thi
 s thesis studies\nfederated machine learning tasks under a number of these
  challenges\, as described below.\n\nA key challenge in cross-device and c
 ross-silo FL is that the data held by individual clients is typically not 
 independent and identically distributed (non-IID). Under this heterogeneit
 y\, each client's model drifts toward its own local minimum\, even when al
 l\nclients participate in the federation. The global model\, aggregated fr
 om these drifted local models\, then performs poorly. A single global mode
 l does have one clear advantage: it can be deployed to clients that were n
 ever seen during training. But because\nno single model suits diverse data
 sets well\, many methods turn to partial model personalization. Relying on
  local parameters\, however\, leaves the server unable to serve new\, unse
 en clients that never joined training. Heterogeneity thus produces both th
 e ‘client-drift’\nproblem and a personalization-generalization trade-o
 ff. A second challenge originates from a practical point that traditional 
 FL tends to ignore: clients are assumed to participate whether or not they
  benefit. This is often unrealistic\, and encouraging participation\nwould
  need a suitable incentive mechanism. FL under client rationality with non
 -monetary incentives remains unstudied and existing algorithms do little t
 o prevent federation collapse in such settings\, especially when data is h
 ighly heterogeneous. A third challenge\narises from the data locality cons
 traint itself. Because data stays strictly on each client\, objectives lik
 e minimax optimization introduce variables tied to individual data points 
 that the server cannot aggregate\, leaving their aggregation undefined. Th
 is\nthesis takes up each of these challenges in turn\, formulating and add
 ressing practical\, non-trivial FL problems that arise from them.\n\nFirst
 \, to address client-drift due to label heterogeneity\, we introduce Adapt
 ive Self-Distillation (ASD)\, a client-side regularizer that adapts to eac
 h client's local dataset using the global model's prediction entropy toget
 her with the client’s label distribution.\nWe justify theoretically that
  ASD reduces client drift and show empirically that it improves generaliza
 tion as it results in the global model converging to a 'flat’ minimum. B
 ecause ASD is plug-and-play\, integrating it into standard FL methods is s
 traightforward\,\nand across multiple real-world benchmarks it leads to su
 bstantial gains. Building on this connection of generalizability to flatne
 ss of loss topology\, we next aim to achieve the generalization objective 
 explicitly through a flatness-constrained optimization\nframework that con
 trols the largest eigenvalue of the Hessian of training-loss. To keep this
  tractable during local training\, we propose FedMAN\, a computationally e
 fficient regularizer that minimizes the per-layer activation norm of the c
 lient-side models.\nWe provide theoretical guarantees for neural nets with
  fully connected and convolutional layers\, showing that minimizing this l
 ayer-wise activation norm bounds the top Hessian eigenvalue of the client'
 s loss\, driving the aggregated model toward a flat minimum.\nIntegrating 
 this regularizer with existing FL methods delivers significant gains in ac
 curacy.\n\nWe now explore how we can adapt a pre-trained model in a federa
 ted setup\, with clients training on local datasets that are unlabeled and
  exhibit significant domain gaps from one client to another. Towards this 
 end\, we address the Federated Source-Free Domain\nAdaptation (FFreeDA) pr
 oblem\, where clients train locally on such unlabeled training samples fro
 m widely heterogeneous data distributions. The client models are initializ
 ed with a pre-trained server model trained on a source domain that is sign
 ificantly different\nfrom every client's local data distribution and also 
 unavailable during the federated training. Adapting Source-Free Domain Ada
 ptation (SFDA) methods to this federated setting struggles with client dri
 ft: the extreme heterogeneity from the domain gaps results\nin unreliable 
 pseudo-labels. We introduce FedSCAl\, an FL framework built on our Server-
 Client Alignment (SCAl) mechanism\, which regularizes client updates by ma
 tching the client and server model predictions on weak and strong augmenta
 tions of the input data\,\nthereby achieving strong server-client alignmen
 t\, reduced client-drift as well as more reliable pseudo labels to train l
 ocal models on. Extensive experiments on benchmark vision datasets show th
 at FedSCAl outperforms the competing FL methods.\n\nIn the works described
  above\, we have trained a global model in which the clients share all mod
 el parameters during the training. However\, to combat data heterogeneity\
 , partial model personalization methods have been proposed. The main drawb
 ack here is that\nthe server model cannot be used for clients which did no
 t participate in the federation. This creates a trade-off in balancing per
 sonalization and generalization. In the context of Vision transformer mode
 ls\, when the training paradigm is prompt tuning\, we\npropose PEP-FedPT\,
  a framework achieving both generalization and personalization in federate
 d prompt tuning of ViTs. At its core is the Class-Contextualized Mixed Pro
 mpt (CCMP) - class-specific prompts maintained alongside a globally shared
  prompt\, which CCMP\nadaptively combines per input using weights from glo
 bal class prototypes and client class priors. This gives per-sample prompt
  personalization without storing any client-dependent trainable parameters
  and the prompts are optimized collaboratively via federated\naveraging. W
 e demonstrate the efficiency of our method on CIFAR-100\, TinyImageNet\, D
 omainNet and iNaturalist datasets.\n\nSo far\, we have assumed that client
 s take part in the federation willingly under server orchestration and wit
 hout conditions. In practice\, this is often not the case. Clients are usu
 ally self-interested and a client may care more about how well the shared\
 nglobal model performs on its held-out data relative to its own local mode
 l performance or any desired threshold. If the clients do not see the bene
 fit they drop out and the whole federation can fall apart. One way to keep
  the clients engaged is to pay them\,\nand such monetary incentives have b
 een studied before. In this work we take a different approach with FedUCA\
 , a non-monetary framework built on a non-payment channel between server a
 nd clients. The server first shares the exact aggregation weights with eac
 h\nclient\, from which the client designs a utility and signals how willin
 g it is to participate. The server then chooses aggregation parameters sat
 isfying each client's individual-rationality (IR) constraint\, sustaining 
 global model performance. Extensive experiments\non standard datasets show
  that\, by prioritizing feasibility of participation\, FedUCA achieves hig
 her client retention and\, in turn\, a stronger global model. As an extens
 ion\, we design FedLoPers\, a lottery-based persuasion scheme that randomi
 zes the information\nshared with clients. Rather than committing to the tr
 ue aggregation weights\, the server shares a distribution over the same. T
 his lottery-inspired design further improves client participation.\n\nIn a
 ll of the above works\, the clients minimize their respective loss functio
 ns. We now consider the federated minimax problems\, a class of problems t
 hat includes distributionally robust optimization and adversarial training
 . Such problems carry an inner maximization\nvariable that is a worst-case
  reweighting of a client's own samples or a perturbation of its own inputs
 \, constructed entirely from local data. Averaging such variables across c
 lients is not just suboptimal but ill-posed. We therefore study the federa
 ted minimax\nproblem in which the minimization variable is shared and aggr
 egated while each client retains a personalized maximization variable that
  is never shared. In the nonconvex-concave regime\, our proposed projected
  personalized ascent scheme attains the standard\nsublinear rate. In the n
 onconvex-PL regime\, the same scheme incurs a stepsize-invariant noise flo
 or from the stochastic ascent updates. To remove it\, we propose VR-SGDA\,
  a variance-reduced ascent step. This eliminates the floor and attains the
  sublinear rate.\nExperiments on subpopulation-robustness and adversarial 
 training confirm our theoretical findings and contributions.\n\n\n\nALL AR
 E WELCOME
CATEGORIES:Events,Ph.D. Thesis Colloquium
END:VEVENT
BEGIN:VTIMEZONE
TZID:Asia/Kolkata
X-LIC-LOCATION:Asia/Kolkata
BEGIN:STANDARD
DTSTART:20251023T110000
TZOFFSETFROM:+0530
TZOFFSETTO:+0530
TZNAME:IST
END:STANDARD
END:VTIMEZONE
END:VCALENDAR