Nathalie Baracaldo

Heshan Fernando, Parikshit Ram, Yi Zhou et al.

Multi-modal learning (MML) aims to integrate information from multiple modalities, which is expected to lead to superior performance over single-modality learning. However, recent studies have shown that MML can underperform, even compared to single-modality approaches, due to imbalanced learning across modalities. Methods have been proposed to alleviate this imbalance issue using different heuristics, which often lead to computationally intensive subroutines. In this paper, we reformulate the MML problem as a multi-objective optimization (MOO) problem that overcomes the imbalanced learning issue among modalities and propose a gradient-based algorithm to solve the modified MML problem. We provide convergence guarantees for the proposed method, and empirical evaluations on popular MML benchmarks showcasing the improved performance of the proposed method over existing balanced MML and MOO baselines, with up to ~20x reduction in subroutine computation time. Our code is available at https://github.com/heshandevaka/MIMO.

Anisa Halimi, Swanand Kadhe, Ambrish Rawat et al.

With privacy legislation empowering the users with the right to be forgotten, it has become essential to make a model amenable for forgetting some of its training data. However, existing unlearning methods in the machine learning context can not be directly applied in the context of distributed settings like federated learning due to the differences in learning protocol and the presence of multiple actors. In this paper, we tackle the problem of federated unlearning for the case of erasing a client by removing the influence of their entire local data from the trained global model. To erase a client, we propose to first perform local unlearning at the client to be erased, and then use the locally unlearned model as the initialization to run very few rounds of federated learning between the server and the remaining clients to obtain the unlearned global model. We empirically evaluate our unlearning method by employing multiple performance measures on three datasets, and demonstrate that our unlearning method achieves comparable performance as the gold standard unlearning method of federated retraining from scratch, while being significantly efficient. Unlike prior works, our unlearning method neither requires global access to the data used for training nor the history of the parameter updates to be stored by the server or any of the clients.

Swanand Ravindra Kadhe, Heiko Ludwig, Nathalie Baracaldo et al.

The effective detection of evidence of financial anomalies requires collaboration among multiple entities who own a diverse set of data, such as a payment network system (PNS) and its partner banks. Trust among these financial institutions is limited by regulation and competition. Federated learning (FL) enables entities to collaboratively train a model when data is either vertically or horizontally partitioned across the entities. However, in real-world financial anomaly detection scenarios, the data is partitioned both vertically and horizontally and hence it is not possible to use existing FL approaches in a plug-and-play manner. Our novel solution, PV4FAD, combines fully homomorphic encryption (HE), secure multi-party computation (SMPC), differential privacy (DP), and randomization techniques to balance privacy and accuracy during training and to prevent inference threats at model deployment time. Our solution provides input privacy through HE and SMPC, and output privacy against inference time attacks through DP. Specifically, we show that, in the honest-but-curious threat model, banks do not learn any sensitive features about PNS transactions, and the PNS does not learn any information about the banks' dataset but only learns prediction labels. We also develop and analyze a DP mechanism to protect output privacy during inference. Our solution generates high-utility models by significantly reducing the per-bank noise level while satisfying distributed DP. To ensure high accuracy, our approach produces an ensemble model, in particular, a random forest. This enables us to take advantage of the well-known properties of ensembles to reduce variance and increase accuracy. Our solution won second prize in the first phase of the U.S. Privacy Enhancing Technologies (PETs) Prize Challenge.

Shuli Jiang, Swanand Ravindra Kadhe, Yi Zhou et al.

The increasing use of large language models (LLMs) trained by third parties raises significant security concerns. In particular, malicious actors can introduce backdoors through poisoning attacks to generate undesirable outputs. While such attacks have been extensively studied in image domains and classification tasks, they remain underexplored for natural language generation (NLG) tasks. To address this gap, we conduct an investigation of various poisoning techniques targeting the LLM's fine-tuning phase via prefix-tuning, a Parameter Efficient Fine-Tuning (PEFT) method. We assess their effectiveness across two generative tasks: text summarization and text completion; and we also introduce new metrics to quantify the success and stealthiness of such NLG poisoning attacks. Through our experiments, we find that the prefix-tuning hyperparameters and trigger designs are the most crucial factors to influence attack success and stealthiness. Moreover, we demonstrate that existing popular defenses are ineffective against our poisoning attacks. Our study presents the first systematic approach to understanding poisoning attacks targeting NLG tasks during fine-tuning via PEFT across a wide range of triggers and attack settings. We hope our findings will aid the AI security community in developing effective defenses against such threats.

Katelinh Jones, Yuya Jeremy Ong, Yi Zhou et al.

Federated Learning (FL) is a paradigm for jointly training machine learning algorithms in a decentralized manner which allows for parties to communicate with an aggregator to create and train a model, without exposing the underlying raw data distribution of the local parties involved in the training process. Most research in FL has been focused on Neural Network-based approaches, however Tree-Based methods, such as XGBoost, have been underexplored in Federated Learning due to the challenges in overcoming the iterative and additive characteristics of the algorithm. Decision tree-based models, in particular XGBoost, can handle non-IID data, which is significant for algorithms used in Federated Learning frameworks since the underlying characteristics of the data are decentralized and have risks of being non-IID by nature. In this paper, we focus on investigating the effects of how Federated XGBoost is impacted by non-IID distributions by performing experiments on various sample size-based data skew scenarios and how these models perform under various non-IID scenarios. We conduct a set of extensive experiments across multiple different datasets and different data skew partitions. Our experimental results demonstrate that despite the various partition ratios, the performance of the models stayed consistent and performed close to or equally well against models that were trained in a centralized manner.

Raj Sanjay Shah, Jing Huang, Keerthiram Murugesan et al.

Unlearning in Large Language Models (LLMs) aims to enhance safety, mitigate biases, and comply with legal mandates, such as the right to be forgotten. However, existing unlearning methods are brittle: minor query modifications, such as multi-hop reasoning and entity aliasing, can recover supposedly forgotten information. As a result, current evaluation metrics often create an illusion of effectiveness, failing to detect these vulnerabilities due to reliance on static, unstructured benchmarks. We propose a dynamic framework that stress tests unlearning robustness using complex structured queries. Our approach first elicits knowledge from the target model (pre-unlearning) and constructs targeted probes, ranging from simple queries to multi-hop chains, allowing precise control over query difficulty. Our experiments show that the framework (1) shows comparable coverage to existing benchmarks by automatically generating semantically equivalent Q&A probes, (2) aligns with prior evaluations, and (3) uncovers new unlearning failures missed by other benchmarks, particularly in multi-hop settings. Furthermore, activation analyses show that single-hop queries typically follow dominant computation pathways, which are more likely to be disrupted by unlearning methods. In contrast, multi-hop queries tend to use alternative pathways that often remain intact, explaining the brittleness of unlearning techniques in multi-hop settings. Our framework enables practical and scalable evaluation of unlearning methods without the need for manual construction of forget test sets, enabling easier adoption for real-world applications. We release the pip package and the code at https://sites.google.com/view/unlearningmirage/home.

Zhexi Lu, Hongliang Chi, Nathalie Baracaldo et al.

Membership inference attacks (MIAs) pose a critical privacy threat to fine-tuned large language models (LLMs), especially when models are adapted to domain-specific tasks using sensitive data. While prior black-box MIA techniques rely on confidence scores or token likelihoods, these signals are often entangled with a sample's intrinsic properties - such as content difficulty or rarity - leading to poor generalization and low signal-to-noise ratios. In this paper, we propose ICP-MIA, a novel MIA framework grounded in the theory of training dynamics, particularly the phenomenon of diminishing returns during optimization. We introduce the Optimization Gap as a fundamental signal of membership: at convergence, member samples exhibit minimal remaining loss-reduction potential, while non-members retain significant potential for further optimization. To estimate this gap in a black-box setting, we propose In-Context Probing (ICP), a training-free method that simulates fine-tuning-like behavior via strategically constructed input contexts. We propose two probing strategies: reference-data-based (using semantically similar public samples) and self-perturbation (via masking or generation). Experiments on three tasks and multiple LLMs show that ICP-MIA significantly outperforms prior black-box MIAs, particularly at low false positive rates. We further analyze how reference data alignment, model type, PEFT configurations, and training schedules affect attack effectiveness. Our findings establish ICP-MIA as a practical and theoretically grounded framework for auditing privacy risks in deployed LLMs.

Momin Abbas, Yi Zhou, Parikshit Ram et al.

In-context learning (ICL) is a new paradigm for natural language processing that utilizes Generative Pre-trained Transformer (GPT)-like models. This approach uses prompts that include in-context demonstrations to generate the corresponding output for a new query input. However, applying ICL in real cases does not scale with the number of samples, and lacks robustness to different prompt templates and demonstration permutations. In this paper, we first show that GPT-like models using ICL result in unreliable predictions based on a new metric based on Shannon entropy. Then, to solve this problem, we propose a new technique called the Linear Probe Calibration (LinC), a method that calibrates the model's output probabilities, resulting in reliable predictions and improved performance, while requiring only minimal additional samples (as few as five labeled data samples). LinC significantly enhances the ICL test performance of GPT models on various benchmark datasets, with an average improvement of up to 21%, and up to a 50% improvement in some cases, and significantly boosts the performance of PEFT methods, especially in the low resource regime. Moreover, LinC achieves lower expected calibration error, and is highly robust to varying label proportions, prompt templates, and demonstration permutations. Our code is available at \url{https://github.com/mominabbass/LinC}.

Jinghan Jia, Nathalie Baracaldo, Sijia Liu

Large reasoning models (LRMs) extend large language models by generating explicit chain-of-thought (CoT) reasoning, significantly improving mathematical and logical problem solving. However, this explicit reasoning process also introduces new safety risks, as unsafe behaviors often emerge within intermediate reasoning trajectories, even when final answers appear harmless. Existing safety alignment approaches primarily rely on supervised fine-tuning (SFT) over safety-oriented long CoT datasets. While intuitive, we find that SFT produces inconsistent safety improvements, degrades reasoning ability, and generalizes poorly across model families. These limitations suggest that purely supervised approaches are insufficient for robust safety alignment in LRMs. To address this, we investigate reinforcement learning (RL) as a complementary optimization framework for LRM safety training. Unlike SFT, RL directly optimizes model policies with reward feedback, enabling more adaptive and stable alignment. Extensive experiments across multiple model families and benchmarks show that RL achieves stronger and more consistent safety gains while maintaining reasoning competence. Further analysis of reflection dynamics and token-level entropy reveals that RL suppresses unsafe exploratory reasoning while preserving reflective depth, leading to safer and more reliable reasoning processes.

Heshan Fernando, Han Shen, Parikshit Ram et al.

The post-training of LLMs, which typically consists of the supervised fine-tuning (SFT) stage and the preference learning stage (RLHF or DPO), is crucial to effective and safe LLM applications. The widely adopted approach in post-training popular open-source LLMs is to sequentially perform SFT and RLHF/DPO. However, this is suboptimal in terms of SFT and RLHF/DPO trade-off: the LLM gradually forgets about the first stage's training when undergoing the second stage's training. This sequential paradigm persists largely due to its simplicity and modularity, which make it easier to implement and manage at scale despite its limitations. We theoretically prove the sub-optimality of sequential post-training and propose a practical joint post-training framework which has theoretical convergence guarantees and empirically outperforms sequential post-training framework, with up to 23% overall performance improvement across multiple LLM evaluation benchmarks, while having minimal computational overhead. Our code is available at https://github.com/heshandevaka/XRIGHT.

Sijia Liu, Yuanshun Yao, Jinghan Jia et al.

We explore machine unlearning (MU) in the domain of large language models (LLMs), referred to as LLM unlearning. This initiative aims to eliminate undesirable data influence (e.g., sensitive or illegal information) and the associated model capabilities, while maintaining the integrity of essential knowledge generation and not affecting causally unrelated information. We envision LLM unlearning becoming a pivotal element in the life-cycle management of LLMs, potentially standing as an essential foundation for developing generative AI that is not only safe, secure, and trustworthy, but also resource-efficient without the need of full retraining. We navigate the unlearning landscape in LLMs from conceptual formulation, methodologies, metrics, and applications. In particular, we highlight the often-overlooked aspects of existing LLM unlearning research, e.g., unlearning scope, data-model interaction, and multifaceted efficacy assessment. We also draw connections between LLM unlearning and related areas such as model editing, influence functions, model explanation, adversarial training, and reinforcement learning. Furthermore, we outline an effective assessment framework for LLM unlearning and explore its applications in copyright and privacy safeguards and sociotechnical harm reduction.

Maria-Irina Nicolae, Mathieu Sinn, Minh Ngoc Tran et al.

Adversarial Robustness Toolbox (ART) is a Python library supporting developers and researchers in defending Machine Learning models (Deep Neural Networks, Gradient Boosted Decision Trees, Support Vector Machines, Random Forests, Logistic Regression, Gaussian Processes, Decision Trees, Scikit-learn Pipelines, etc.) against adversarial threats and helps making AI systems more secure and trustworthy. Machine Learning models are vulnerable to adversarial examples, which are inputs (images, texts, tabular data, etc.) deliberately modified to produce a desired response by the Machine Learning model. ART provides the tools to build and deploy defences and test them with adversarial attacks. Defending Machine Learning models involves certifying and verifying model robustness and model hardening with approaches such as pre-processing inputs, augmenting training data with adversarial samples, and leveraging runtime detection methods to flag any inputs that might have been modified by an adversary. The attacks implemented in ART allow creating adversarial attacks against Machine Learning models which is required to test defenses with state-of-the-art threat models. Supported Machine Learning Libraries include TensorFlow (v1 and v2), Keras, PyTorch, MXNet, Scikit-learn, XGBoost, LightGBM, CatBoost, and GPy. The source code of ART is released with MIT license at https://github.com/IBM/adversarial-robustness-toolbox. The release includes code examples, notebooks with tutorials and documentation (http://adversarial-robustness-toolbox.readthedocs.io).

Woogyeol Jin, Taywon Min, Yongjin Yang et al.

On-policy distillation is a promising approach for transferring knowledge between language models, where a student learns from dense token-level signals along its own trajectories. This framework typically uses reverse KL divergence, encouraging the student to match the teacher's high-confidence predictions. However, we show that the mode-seeking property of reverse KL reduces generation diversity and yields unstable learning signals when the teacher distribution has high entropy. To address this, we introduce Entropy-Aware On-Policy Distillation. Our key idea is augmenting the standard reverse KL objective with forward KL when teacher entropy is high, capturing the full range of plausible outputs while retaining precise imitation elsewhere. It balances mode-seeking precision with mode-covering robustness without sacrificing on-policy training efficiency. Experiments show that our method maintains generation diversity (sustained token-level entropy) and improves student-teacher alignment (lower forward KL on high-entropy tokens). Across six math reasoning benchmarks, this yields Pass@8 accuracy gains of +1.37 for Qwen3-0.6B-Base, +2.39 for Qwen3-1.7B-Base, and +5.05 for Qwen3-4B-Base compared to baseline on-policy distillation methods. These results demonstrate that accounting for teacher uncertainty is essential for maintaining diversity and achieving effective knowledge transfer.

Jinghan Jia, Jiancheng Liu, Yihua Zhang et al.

The need for effective unlearning mechanisms in large language models (LLMs) is increasingly urgent, driven by the necessity to adhere to data regulations and foster ethical generative AI practices. Despite growing interest of LLM unlearning, much of the existing research has focused on varied unlearning method designs to boost effectiveness and efficiency. However, the inherent relationship between model weights and LLM unlearning has not been extensively examined. In this paper, we systematically explore how model weights interact with unlearning processes in LLMs and we design the weight attribution-guided LLM unlearning method, WAGLE, which unveils the interconnections between 'influence' of weights and 'influence' of data to forget and retain in LLM generation. By strategically guiding the LLM unlearning across different types of unlearning methods and tasks, WAGLE can erase the undesired content, while maintaining the performance of the original tasks. We refer to the weight attribution-guided LLM unlearning method as WAGLE, which unveils the interconnections between 'influence' of weights and 'influence' of data to forget and retain in LLM generation. Our extensive experiments show that WAGLE boosts unlearning performance across a range of LLM unlearning methods such as gradient difference and (negative) preference optimization, applications such as fictitious unlearning, malicious use prevention, and copyrighted information removal, and models including Zephyr-7b-beta and Llama2-7b. To the best of our knowledge, our work offers the first principled method for attributing and pinpointing the influential weights in enhancing LLM unlearning. It stands in contrast to previous methods that lack weight attribution and simpler weight attribution techniques.

Shuli Jiang, Swanand Ravindra Kadhe, Yi Zhou et al.

Growing applications of large language models (LLMs) trained by a third party raise serious concerns on the security vulnerability of LLMs.It has been demonstrated that malicious actors can covertly exploit these vulnerabilities in LLMs through poisoning attacks aimed at generating undesirable outputs. While poisoning attacks have received significant attention in the image domain (e.g., object detection), and classification tasks, their implications for generative models, particularly in the realm of natural language generation (NLG) tasks, remain poorly understood. To bridge this gap, we perform a comprehensive exploration of various poisoning techniques to assess their effectiveness across a range of generative tasks. Furthermore, we introduce a range of metrics designed to quantify the success and stealthiness of poisoning attacks specifically tailored to NLG tasks. Through extensive experiments on multiple NLG tasks, LLMs and datasets, we show that it is possible to successfully poison an LLM during the fine-tuning stage using as little as 1\% of the total tuning data samples. Our paper presents the first systematic approach to comprehend poisoning attacks targeting NLG tasks considering a wide range of triggers and attack settings. We hope our findings will assist the AI security community in devising appropriate defenses against such threats.

Swanand Ravindra Kadhe, Anisa Halimi, Ambrish Rawat et al.

Training large language models (LLMs) is a costly endeavour in terms of time and computational resources. The large amount of training data used during the unsupervised pre-training phase makes it difficult to verify all data and, unfortunately, undesirable data may be ingested during training. Re-training from scratch is impractical and has led to the creation of the 'unlearning' discipline where models are modified to "unlearn" undesirable information without retraining. However, any modification can alter the behaviour of LLMs, especially on key dimensions such as fairness. This is the first work that examines this interplay between unlearning and fairness for LLMs. In particular, we focus on a popular unlearning framework known as SISA [Bourtoule et al., 2021], which creates an ensemble of models trained on disjoint shards. We evaluate the performance-fairness trade-off for SISA, and empirically demsontrate that SISA can indeed reduce fairness in LLMs. To remedy this, we propose post-processing bias mitigation techniques for ensemble models produced by SISA. We adapt the post-processing fairness improvement technique from [Hardt et al., 2016] to design three methods that can handle model ensembles, and prove that one of the methods is an optimal fair predictor for ensemble of models. Through experimental results, we demonstrate the efficacy of our post-processing framework called 'FairSISA'.

Xinran Wang, Qi Le, Ammar Ahmed et al.

Ensuring that generative AI systems align with human values is essential but challenging, especially when considering multiple human values and their potential trade-offs. Since human values can be personalized and dynamically change over time, the desirable levels of value alignment vary across different ethnic groups, industry sectors, and user cohorts. Within existing frameworks, it is hard to define human values and align AI systems accordingly across different directions simultaneously, such as harmlessness, helpfulness, and positiveness. To address this, we develop a novel, first-principle approach called Multi-Human-Value Alignment Palette (MAP), which navigates the alignment across multiple human values in a structured and reliable way. MAP formulates the alignment problem as an optimization task with user-defined constraints, which define human value targets. It can be efficiently solved via a primal-dual approach, which determines whether a user-defined alignment target is achievable and how to achieve it. We conduct a detailed theoretical analysis of MAP by quantifying the trade-offs between values, the sensitivity to constraints, the fundamental connection between multi-value alignment and sequential alignment, and proving that linear weighted rewards are sufficient for multi-value alignment. Extensive experiments demonstrate MAP's ability to align multiple values in a principled manner while delivering strong empirical performance across various tasks.

Changsheng Wang, Chongyu Fan, Yihua Zhang et al.

Recent advances in large reasoning models (LRMs) have enabled strong chain-of-thought (CoT) generation through test-time computation. While these multi-step reasoning capabilities represent a major milestone in language model performance, they also introduce new safety risks. In this work, we present the first systematic study to revisit the problem of machine unlearning in the context of LRMs. Machine unlearning refers to the process of removing the influence of sensitive, harmful, or undesired data or knowledge from a trained model without full retraining. We show that conventional unlearning algorithms, originally designed for non-reasoning models, are inadequate for LRMs. In particular, even when final answers are successfully erased, sensitive information often persists within the intermediate reasoning steps, i.e., CoT trajectories. To address this challenge, we extend conventional unlearning and propose Reasoning-aware Representation Misdirection for Unlearning ($R^2MU$), a novel method that effectively suppresses sensitive reasoning traces and prevents the generation of associated final answers, while preserving the model's reasoning ability. Our experiments demonstrate that $R^2MU$ significantly reduces sensitive information leakage within reasoning traces and achieves strong performance across both safety and reasoning benchmarks, evaluated on state-of-the-art models such as DeepSeek-R1-Distill-LLaMA-8B and DeepSeek-R1-Distill-Qwen-14B.

Changsheng Wang, Yihua Zhang, Jinghan Jia et al.

Machine unlearning offers a promising solution to privacy and safety concerns in large language models (LLMs) by selectively removing targeted knowledge while preserving utility. However, current methods are highly sensitive to downstream fine-tuning, which can quickly recover forgotten information-even from unrelated tasks. To address this, we introduce invariance into unlearning for the first time, inspired by invariant risk minimization (IRM). Building on this principle, we propose invariant LLM unlearning (ILU), a regularization-based framework that enhances robustness. Notably, ILU generalizes well to diverse fine-tuning tasks, even when trained using a single dataset. A task vector analysis is also provided to further elucidate the rationale behind ILU's effectiveness. Extensive experiments on the WMDP and MUSE benchmark, reveal that ILU significantly outperforms state-of-the-art unlearning methods, including negative preference optimization (NPO) and representation misdirection for unlearning (RMU). Notably, ILU achieves superior unlearning robustness across diverse downstream fine-tuning scenarios (e.g., math, paraphrase detection, and sentiment analysis) while preserving the fine-tuning performance.

Yuetian Chen, Zhiqi Wang, Nathalie Baracaldo et al.

Membership inference attacks (MIAs) pose a critical threat to the privacy of training data in deep learning. Despite significant progress in attack methodologies, our understanding of when and how models encode membership information during training remains limited. This paper presents a dynamic analytical framework for dissecting and quantifying privacy leakage dynamics at the individual sample level. By tracking per-sample vulnerabilities on an FPR-TPR plane throughout training, our framework systematically measures how factors such as dataset complexity, model architecture, and optimizer choice influence the rate and severity at which samples become vulnerable. Crucially, we discover a robust correlation between a sample's intrinsic learning difficulty, and find that the privacy risk of samples highly vulnerable in the final trained model is largely determined early during training. Our results thus provide a deeper understanding of how privacy risks dynamically emerge during training, laying the groundwork for proactive, privacy-aware model training strategies.

Zhiqi Wang, Chengyu Zhang, Yuetian Chen et al.

Membership inference attacks (MIAs) pose a significant threat to the privacy of machine learning models and are widely used as tools for privacy assessment, auditing, and machine unlearning. While prior MIA research has primarily focused on performance metrics such as AUC, accuracy, and TPR@low FPR - either by developing new methods to enhance these metrics or using them to evaluate privacy solutions - we found that it overlooks the disparities among different attacks. These disparities, both between distinct attack methods and between multiple instantiations of the same method, have crucial implications for the reliability and completeness of MIAs as privacy evaluation tools. In this paper, we systematically investigate these disparities through a novel framework based on coverage and stability analysis. Extensive experiments reveal significant disparities in MIAs, their potential causes, and their broader implications for privacy evaluation. To address these challenges, we propose an ensemble framework with three distinct strategies to harness the strengths of state-of-the-art MIAs while accounting for their disparities. This framework not only enables the construction of more powerful attacks but also provides a more robust and comprehensive methodology for privacy evaluation.

Jinghan Jia, Hadi Reisizadeh, Chongyu Fan et al.

Large language models (LLMs) have shown remarkable reasoning capabilities when trained with chain-of-thought (CoT) supervision. However, the long and verbose CoT traces, especially those distilled from large reasoning models (LRMs) such as DeepSeek-R1, significantly increase training costs during the distillation process, where a non-reasoning base model is taught to replicate the reasoning behavior of an LRM. In this work, we study the problem of CoT condensation for resource-efficient reasoning training, aimed at pruning intermediate reasoning steps (i.e., thoughts) in CoT traces, enabling supervised model training on length-reduced CoT data while preserving both answer accuracy and the model's ability to generate coherent reasoning. Our rationale is that CoT traces typically follow a three-stage structure: problem understanding, exploration, and solution convergence. Through empirical analysis, we find that retaining the structure of the reasoning trace, especially the early stage of problem understanding (rich in reflective cues) and the final stage of solution convergence, is sufficient to achieve lossless reasoning supervision. To this end, we propose an Edge-Preserving Condensation method, EPiC, which selectively retains only the initial and final segments of each CoT trace while discarding the middle portion. This design draws an analogy to preserving the "edge" of a reasoning trajectory, capturing both the initial problem framing and the final answer synthesis, to maintain logical continuity. Experiments across multiple model families (Qwen and LLaMA) and benchmarks show that EPiC reduces training time by over 34% while achieving lossless reasoning accuracy on MATH500, comparable to full CoT supervision. To the best of our knowledge, this is the first study to explore thought-level CoT condensation for efficient reasoning model distillation.

Swanand Ravindra Kadhe, Farhan Ahmed, Dennis Wei et al.

Large language models (LLMs) have shown to pose social and ethical risks such as generating toxic language or facilitating malicious use of hazardous knowledge. Machine unlearning is a promising approach to improve LLM safety by directly removing harmful behaviors and knowledge. In this paper, we propose "SPlit, UNlearn, MerGE" (SPUNGE), a framework that can be used with any unlearning method to amplify its effectiveness. SPUNGE leverages data attributes during unlearning by splitting unlearning data into subsets based on specific attribute values, unlearning each subset separately, and merging the unlearned models. We empirically demonstrate that SPUNGE significantly improves the performance of two recent unlearning methods on state-of-the-art LLMs while maintaining their general capabilities on standard academic benchmarks.

Timothy Castiglia, Yi Zhou, Shiqiang Wang et al.

We propose LESS-VFL, a communication-efficient feature selection method for distributed systems with vertically partitioned data. We consider a system of a server and several parties with local datasets that share a sample ID space but have different feature sets. The parties wish to collaboratively train a model for a prediction task. As part of the training, the parties wish to remove unimportant features in the system to improve generalization, efficiency, and explainability. In LESS-VFL, after a short pre-training period, the server optimizes its part of the global model to determine the relevant outputs from party models. This information is shared with the parties to then allow local feature selection without communication. We analytically prove that LESS-VFL removes spurious features from model training. We provide extensive empirical evidence that LESS-VFL can achieve high accuracy and remove spurious features at a fraction of the communication cost of other feature selection approaches.

Nathalie Baracaldo, Ali Anwar, Mark Purcell et al.

Federated Learning (FL) is a novel paradigm for the shared training of models based on decentralized and private data. With respect to ethical guidelines, FL is promising regarding privacy, but needs to excel vis-à-vis transparency and trustworthiness. In particular, FL has to address the accountability of the parties involved and their adherence to rules, law and principles. We introduce AF^2 Framework, where we instrument FL with accountability by fusing verifiable claims with tamper-evident facts, into reproducible arguments. We build on AI FactSheets for instilling transparency and trustworthiness into the AI lifecycle and expand it to incorporate dynamic and nested facts, as well as complex model compositions in FL. Based on our approach, an auditor can validate, reproduce and certify a FL process. This can be directly applied in practice to address the challenges of AI engineering and ethics.

Yi Zhou, Parikshit Ram, Theodoros Salonidis et al.

We address the relatively unexplored problem of hyper-parameter optimization (HPO) for federated learning (FL-HPO). We introduce Federated Loss SuRface Aggregation (FLoRA), a general FL-HPO solution framework that can address use cases of tabular data and any Machine Learning (ML) model including gradient boosting training algorithms and therefore further expands the scope of FL-HPO. FLoRA enables single-shot FL-HPO: identifying a single set of good hyper-parameters that are subsequently used in a single FL training. Thus, it enables FL-HPO solutions with minimal additional communication overhead compared to FL training without HPO. We theoretically characterize the optimality gap of FL-HPO, which explicitly accounts for the heterogeneous non-IID nature of the parties' local data distributions, a dominant characteristic of FL systems. Our empirical evaluation of FLoRA for multiple ML algorithms on seven OpenML datasets demonstrates significant model accuracy improvements over the considered baseline, and robustness to increasing number of parties involved in FL-HPO training.

Yi Zhou, Parikshit Ram, Theodoros Salonidis et al.

We address the relatively unexplored problem of hyper-parameter optimization (HPO) for federated learning (FL-HPO). We introduce Federated Loss suRface Aggregation (FLoRA), the first FL-HPO solution framework that can address use cases of tabular data and gradient boosting training algorithms in addition to stochastic gradient descent/neural networks commonly addressed in the FL literature. The framework enables single-shot FL-HPO, by first identifying a good set of hyper-parameters that are used in a **single** FL training. Thus, it enables FL-HPO solutions with minimal additional communication overhead compared to FL training without HPO. Our empirical evaluation of FLoRA for Gradient Boosted Decision Trees on seven OpenML data sets demonstrates significant model accuracy improvements over the considered baseline, and robustness to increasing number of parties involved in FL-HPO training.

Runhua Xu, Nathalie Baracaldo, James Joshi

Machine learning (ML) is increasingly being adopted in a wide variety of application domains. Usually, a well-performing ML model relies on a large volume of training data and high-powered computational resources. Such a need for and the use of huge volumes of data raise serious privacy concerns because of the potential risks of leakage of highly privacy-sensitive information; further, the evolving regulatory environments that increasingly restrict access to and use of privacy-sensitive data add significant challenges to fully benefiting from the power of ML for data-driven applications. A trained ML model may also be vulnerable to adversarial attacks such as membership, attribute, or property inference attacks and model inversion attacks. Hence, well-designed privacy-preserving ML (PPML) solutions are critically needed for many emerging applications. Increasingly, significant research efforts from both academia and industry can be seen in PPML areas that aim toward integrating privacy-preserving techniques into ML pipeline or specific algorithms, or designing various PPML architectures. In particular, existing PPML research cross-cut ML, systems and applications design, as well as security and privacy areas; hence, there is a critical need to understand state-of-the-art research, related challenges and a research roadmap for future research in PPML area. In this paper, we systematically review and summarize existing privacy-preserving approaches and propose a Phase, Guarantee, and Utility (PGU) triad based model to understand and guide the evaluation of various PPML solutions by decomposing their privacy-preserving functionalities. We discuss the unique characteristics and challenges of PPML and outline possible research directions that leverage as well as benefit multiple research communities such as ML, distributed systems, security and privacy.

Kamala Varma, Yi Zhou, Nathalie Baracaldo et al.

Federated learning has arisen as a mechanism to allow multiple participants to collaboratively train a model without sharing their data. In these settings, participants (workers) may not trust each other fully; for instance, a set of competitors may collaboratively train a machine learning model to detect fraud. The workers provide local gradients that a central server uses to update a global model. This global model can be corrupted when Byzantine workers send malicious gradients, which necessitates robust methods for aggregating gradients that mitigate the adverse effects of Byzantine inputs. Existing robust aggregation algorithms are often computationally expensive and only effective under strict assumptions. In this paper, we introduce LayerwisE Gradient AggregatTiOn (LEGATO), an aggregation algorithm that is, by contrast, scalable and generalizable. Informed by a study of layer-specific responses of gradients to Byzantine attacks, LEGATO employs a dynamic gradient reweighing scheme that is novel in its treatment of gradients based on layer-specific robustness. We show that LEGATO is more computationally efficient than multiple state-of-the-art techniques and more generally robust across a variety of attack settings in practice. We also demonstrate LEGATO's benefits for gradient descent convergence in the absence of an attack.

Runhua Xu, Nathalie Baracaldo, Yi Zhou et al.

Federated learning (FL) has been proposed to allow collaborative training of machine learning (ML) models among multiple parties where each party can keep its data private. In this paradigm, only model updates, such as model weights or gradients, are shared. Many existing approaches have focused on horizontal FL, where each party has the entire feature set and labels in the training data set. However, many real scenarios follow a vertically-partitioned FL setup, where a complete feature set is formed only when all the datasets from the parties are combined, and the labels are only available to a single party. Privacy-preserving vertical FL is challenging because complete sets of labels and features are not owned by one entity. Existing approaches for vertical FL require multiple peer-to-peer communications among parties, leading to lengthy training times, and are restricted to (approximated) linear models and just two parties. To close this gap, we propose FedV, a framework for secure gradient computation in vertical settings for several widely used ML models such as linear models, logistic regression, and support vector machines. FedV removes the need for peer-to-peer communication among parties by using functional encryption schemes; this allows FedV to achieve faster training times. It also works for larger and changing sets of parties. We empirically demonstrate the applicability for multiple types of ML models and show a reduction of 10%-70% of training time and 80% to 90% in data transfer with respect to the state-of-the-art approaches.

Syed Zawad, Ahsan Ali, Pin-Yu Chen et al.

Data heterogeneity has been identified as one of the key features in federated learning but often overlooked in the lens of robustness to adversarial attacks. This paper focuses on characterizing and understanding its impact on backdooring attacks in federated learning through comprehensive experiments using synthetic and the LEAF benchmarks. The initial impression driven by our experimental results suggests that data heterogeneity is the dominant factor in the effectiveness of attacks and it may be a redemption for defending against backdooring as it makes the attack less efficient, more challenging to design effective attack strategies, and the attack result also becomes less predictable. However, with further investigations, we found data heterogeneity is more of a curse than a redemption as the attack effectiveness can be significantly boosted by simply adjusting the client-side backdooring timing. More importantly,data heterogeneity may result in overfitting at the local training of benign clients, which can be utilized by attackers to disguise themselves and fool skewed-feature based defenses. In addition, effective attack strategies can be made by adjusting attack data distribution. Finally, we discuss the potential directions of defending the curses brought by data heterogeneity. The results and lessons learned from our extensive experiments and analysis offer new insights for designing robust federated learning methods and systems

Yuya Jeremy Ong, Yi Zhou, Nathalie Baracaldo et al.

Federated Learning (FL) is an approach to collaboratively train a model across multiple parties without sharing data between parties or an aggregator. It is used both in the consumer domain to protect personal data as well as in enterprise settings, where dealing with data domicile regulation and the pragmatics of data silos are the main drivers. While gradient boosted tree implementations such as XGBoost have been very successful for many use cases, its federated learning adaptations tend to be very slow due to using cryptographic and privacy methods and have not experienced widespread use. We propose the Party-Adaptive XGBoost (PAX) for federated learning, a novel implementation of gradient boosting which utilizes a party adaptive histogram aggregation method, without the need for data encryption. It constructs a surrogate representation of the data distribution for finding splits of the decision tree. Our experimental results demonstrate strong model performance, especially on non-IID distributions, and significantly faster training run-time across different data sets than existing federated implementations. This approach makes the use of gradient boosted trees practical in enterprise federated learning.

Annie Abay, Yi Zhou, Nathalie Baracaldo et al.

As methods to create discrimination-aware models develop, they focus on centralized ML, leaving federated learning (FL) unexplored. FL is a rising approach for collaborative ML, in which an aggregator orchestrates multiple parties to train a global model without sharing their training data. In this paper, we discuss causes of bias in FL and propose three pre-processing and in-processing methods to mitigate bias, without compromising data privacy, a key FL requirement. As data heterogeneity among parties is one of the challenging characteristics of FL, we conduct experiments over several data distributions to analyze their effects on model performance, fairness metrics, and bias learning patterns. We conduct a comprehensive analysis of our proposed techniques, the results demonstrating that these methods are effective even when parties have skewed data distributions or as little as 20% of parties employ the methods.

Heiko Ludwig, Nathalie Baracaldo, Gegi Thomas et al.

Federated Learning (FL) is an approach to conduct machine learning without centralizing training data in a single place, for reasons of privacy, confidentiality or data volume. However, solving federated machine learning problems raises issues above and beyond those of centralized machine learning. These issues include setting up communication infrastructure between parties, coordinating the learning process, integrating party results, understanding the characteristics of the training data sets of different participating parties, handling data heterogeneity, and operating with the absence of a verification data set. IBM Federated Learning provides infrastructure and coordination for federated learning. Data scientists can design and run federated learning jobs based on existing, centralized machine learning models and can provide high-level instructions on how to run the federation. The framework applies to both Deep Neural Networks as well as ``traditional'' approaches for the most common machine learning libraries. {\proj} enables data scientists to expand their scope from centralized to federated machine learning, minimizing the learning curve at the outset while also providing the flexibility to deploy to different compute environments and design custom fusion algorithms.

Zheng Chai, Ahsan Ali, Syed Zawad et al.

Federated Learning (FL) enables learning a shared model across many clients without violating the privacy requirements. One of the key attributes in FL is the heterogeneity that exists in both resource and data due to the differences in computation and communication capacity, as well as the quantity and content of data among different clients. We conduct a case study to show that heterogeneity in resource and data has a significant impact on training time and model accuracy in conventional FL systems. To this end, we propose TiFL, a Tier-based Federated Learning System, which divides clients into tiers based on their training performance and selects clients from the same tier in each training round to mitigate the straggler problem caused by heterogeneity in resource and data quantity. To further tame the heterogeneity caused by non-IID (Independent and Identical Distribution) data and resources, TiFL employs an adaptive tier selection approach to update the tiering on-the-fly based on the observed training performance and accuracy overtime. We prototype TiFL in a FL testbed following Google's FL architecture and evaluate it using popular benchmarks and the state-of-the-art FL benchmark LEAF. Experimental evaluation shows that TiFL outperforms the conventional FL in various heterogeneous conditions. With the proposed adaptive tier selection policy, we demonstrate that TiFL achieves much faster training performance while keeping the same (and in some cases - better) test accuracy across the board.

Runhua Xu, Nathalie Baracaldo, Yi Zhou et al.

Federated learning has emerged as a promising approach for collaborative and privacy-preserving learning. Participants in a federated learning process cooperatively train a model by exchanging model parameters instead of the actual training data, which they might want to keep private. However, parameter interaction and the resulting model still might disclose information about the training data used. To address these privacy concerns, several approaches have been proposed based on differential privacy and secure multiparty computation (SMC), among others. They often result in large communication overhead and slow training time. In this paper, we propose HybridAlpha, an approach for privacy-preserving federated learning employing an SMC protocol based on functional encryption. This protocol is simple, efficient and resilient to participants dropping out. We evaluate our approach regarding the training time and data volume exchanged using a federated learning process to train a CNN on the MNIST data set. Evaluation against existing crypto-based SMC solutions shows that HybridAlpha can reduce the training time by 68% and data transfer volume by 92% on average while providing the same model performance and privacy guarantees as the existing solutions.

Stacey Truex, Nathalie Baracaldo, Ali Anwar et al.

Federated learning facilitates the collaborative training of models without the sharing of raw data. However, recent attacks demonstrate that simply maintaining data locality during training processes does not provide sufficient privacy guarantees. Rather, we need a federated learning system capable of preventing inference over both the messages exchanged during training and the final trained model while ensuring the resulting model also has acceptable predictive accuracy. Existing federated learning approaches either use secure multiparty computation (SMC) which is vulnerable to inference or differential privacy which can lead to low accuracy given a large number of parties with relatively small amounts of data each. In this paper, we present an alternative approach that utilizes both differential privacy and SMC to balance these trade-offs. Combining differential privacy with secure multiparty computation enables us to reduce the growth of noise injection as the number of parties increases without sacrificing privacy while maintaining a pre-defined rate of trust. Our system is therefore a scalable approach that protects against inference threats and produces models with high accuracy. Additionally, our system can be used to train a variety of machine learning models, which we validate with experimental results on 3 different machine learning algorithms. Our experiments demonstrate that our approach out-performs state of the art solutions.

Bryant Chen, Wilka Carvalho, Nathalie Baracaldo et al.

While machine learning (ML) models are being increasingly trusted to make decisions in different and varying areas, the safety of systems using such models has become an increasing concern. In particular, ML models are often trained on data from potentially untrustworthy sources, providing adversaries with the opportunity to manipulate them by inserting carefully crafted samples into the training set. Recent work has shown that this type of attack, called a poisoning attack, allows adversaries to insert backdoors or trojans into the model, enabling malicious behavior with simple external backdoor triggers at inference time and only a blackbox perspective of the model itself. Detecting this type of attack is challenging because the unexpected behavior occurs only when a backdoor trigger, which is known only to the adversary, is present. Model users, either direct users of training data or users of pre-trained model from a catalog, may not guarantee the safe operation of their ML-based system. In this paper, we propose a novel approach to backdoor detection and removal for neural networks. Through extensive experimental results, we demonstrate its effectiveness for neural networks classifying text and images. To the best of our knowledge, this is the first methodology capable of detecting poisonous data crafted to insert backdoors and repairing the model that does not require a verified and trusted dataset.