arXiv subjects

Ilias Leontiadis

Publications and source records attributed to Ilias Leontiadis.

At least 19 recordsLinked to original sources

arXiv2026

Rethinking Rubric Generation for Improving LLM Judge and Reward Modeling for Open-ended Tasks

Recently, rubrics have been used to guide LLM judges in capturing subjective, nuanced, multi-dimensional human preferences, and have been extended from evaluation to reward signals for reinforcement fine-tuning (RFT). However, rubric generation remains hard to control: rubrics often lack coverage, conflate dimensions, misalign preference direction, and contain redundant or highly correlated criteria, degrading judge accuracy and producing suboptimal rewards during RFT. We propose RRD, a principled framework for rubric refinement built on a recursive decompose-filter cycle. RRD decomposes coarse rubrics into fine-grained, discriminative criteria, expanding coverage while sharpening separation between responses. A complementary filtering mechanism removes misaligned and redundant rubrics, and a correlation-aware weighting scheme prevents over-representing highly correlated criteria, yielding rubric sets that are informative, comprehensive, and non-redundant. Empirically, RRD delivers large, consistent gains across both evaluation and training: it improves preference-judgment accuracy on JudgeBench and PPE for both GPT-4o and Llama3.1-405B judges, achieving top performance in all settings with up to +17.7 points on JudgeBench. When used as the reward source for RFT on WildChat, it yields substantially stronger and more stable learning signals, boosting reward by up to 160% (Qwen3-4B) and 60% (Llama3.1-8B) versus 10-20% for prior rubric baselines, with gains that transfer to HealthBench-Hard and BiGGen Bench. Overall, RRD establishes recursive rubric refinement as a scalable and interpretable foundation for LLM judging and reward modeling in open-ended domains.

cs.LG↗

arXiv2026

Balanced Accuracy: The Right Metric for Evaluating LLM Judges -- Explained through Youden's J statistic

Rigorous evaluation of large language models (LLMs) relies on comparing models by the prevalence of desirable or undesirable behaviors, such as task pass rates or policy violations. These prevalence estimates are produced by a classifier, either an LLM-as-a-judge or human annotators, making the choice of classifier central to trustworthy evaluation. Common metrics used for this choice, such as Accuracy, Precision, and F1, are sensitive to class imbalance and to arbitrary choices of positive class, and can favor judges that distort prevalence estimates. We show that Youden's $J$ statistic is theoretically aligned with choosing the best judge to compare models, and that Balanced Accuracy is an equivalent linear transformation of $J$. Through both analytical arguments and empirical examples and simulations, we demonstrate how selecting judges using Balanced Accuracy leads to better, more robust classifier selection.

cs.LG↗

arXiv2025

Training AI Co-Scientists Using Rubric Rewards

AI co-scientists are emerging as a tool to assist human researchers in achieving their research goals. A crucial feature of these AI co-scientists is the ability to generate a research plan given a set of aims and constraints. The plan may be used by researchers for brainstorming, or may even be implemented after further refinement. However, language models currently struggle to generate research plans that follow all constraints and implicit requirements. In this work, we study how to leverage the vast corpus of existing research papers to train language models that generate better research plans. We build a scalable, diverse training corpus by automatically extracting research goals and goal-specific grading rubrics from papers across several domains. We then train models for research plan generation via reinforcement learning with self-grading. A frozen copy of the initial policy acts as the grader during training, with the rubrics creating a generator-verifier gap that enables improvements without external human supervision. To validate this approach, we conduct a study with human experts for machine learning research goals, spanning 225 hours. The experts prefer plans generated by our finetuned Qwen3-30B-A3B model over the initial model for 70% of research goals, and approve 84% of the automatically extracted goal-specific grading rubrics. To assess generality, we also extend our approach to research goals from medical papers, and new arXiv preprints, evaluating with a jury of frontier models. Our finetuning yields 12-22% relative improvements and significant cross-domain generalization, proving effective even in problem settings like medical research where execution feedback is infeasible. Together, these findings demonstrate the potential of a scalable, automated training recipe as a step towards improving general AI co-scientists.

cs.LG↗

arXiv2024

Evaluating Privacy Leakage in Split Learning

Privacy-Preserving machine learning (PPML) can help us train and deploy models that utilize private information. In particular, on-device machine learning allows us to avoid sharing raw data with a third-party server during inference. On-device models are typically less accurate when compared to their server counterparts due to the fact that (1) they typically only rely on a small set of on-device features and (2) they need to be small enough to run efficiently on end-user devices. Split Learning (SL) is a promising approach that can overcome these limitations. In SL, a large machine learning model is divided into two parts, with the bigger part residing on the server side and a smaller part executing on-device, aiming to incorporate the private features. However, end-to-end training of such models requires exchanging gradients at the cut layer, which might encode private features or labels. In this paper, we provide insights into potential privacy risks associated with SL. Furthermore, we also investigate the effectiveness of various mitigation strategies. Our results indicate that the gradients significantly improve the attackers' effectiveness in all tested datasets reaching almost perfect reconstruction accuracy for some features. However, a small amount of differential privacy (DP) can effectively mitigate this risk without causing significant training degradation.

cs.LG↗

arXiv2023

GPU-based Private Information Retrieval for On-Device Machine Learning Inference

On-device machine learning (ML) inference can enable the use of private user data on user devices without revealing them to remote servers. However, a pure on-device solution to private ML inference is impractical for many applications that rely on embedding tables that are too large to be stored on-device. In particular, recommendation models typically use multiple embedding tables each on the order of 1-10 GBs of data, making them impractical to store on-device. To overcome this barrier, we propose the use of private information retrieval (PIR) to efficiently and privately retrieve embeddings from servers without sharing any private information. As off-the-shelf PIR algorithms are usually too computationally intensive to directly use for latency-sensitive inference tasks, we 1) propose novel GPU-based acceleration of PIR, and 2) co-design PIR with the downstream ML application to obtain further speedup. Our GPU acceleration strategy improves system throughput by more than $20 \times$ over an optimized CPU PIR implementation, and our PIR-ML co-design provides an over $5 \times$ additional throughput improvement at fixed model quality. Together, for various on-device ML applications such as recommendation and language modeling, our system on a single V100 GPU can serve up to $100,000$ queries per second -- a $>100 \times$ throughput improvement over a CPU-based baseline -- while maintaining model accuracy.

cs.CR↗

arXiv2022

FEL: High Capacity Learning for Recommendation and Ranking via Federated Ensemble Learning

Federated learning (FL) has emerged as an effective approach to address consumer privacy needs. FL has been successfully applied to certain machine learning tasks, such as training smart keyboard models and keyword spotting. Despite FL's initial success, many important deep learning use cases, such as ranking and recommendation tasks, have been limited from on-device learning. One of the key challenges faced by practical FL adoption for DL-based ranking and recommendation is the prohibitive resource requirements that cannot be satisfied by modern mobile systems. We propose Federated Ensemble Learning (FEL) as a solution to tackle the large memory requirement of deep learning ranking and recommendation tasks. FEL enables large-scale ranking and recommendation model training on-device by simultaneously training multiple model versions on disjoint clusters of client devices. FEL integrates the trained sub-models via an over-arch layer into an ensemble model that is hosted on the server. Our experiments demonstrate that FEL leads to 0.43-2.31% model quality improvement over traditional on-device federated learning - a significant improvement for ranking and recommendation system use cases.

cs.IR↗

arXiv2022

FjORD: Fair and Accurate Federated Learning under heterogeneous targets with Ordered Dropout

Federated Learning (FL) has been gaining significant traction across different ML tasks, ranging from vision to keyboard predictions. In large-scale deployments, client heterogeneity is a fact and constitutes a primary problem for fairness, training performance and accuracy. Although significant efforts have been made into tackling statistical data heterogeneity, the diversity in the processing capabilities and network bandwidth of clients, termed as system heterogeneity, has remained largely unexplored. Current solutions either disregard a large portion of available devices or set a uniform limit on the model's capacity, restricted by the least capable participants. In this work, we introduce Ordered Dropout, a mechanism that achieves an ordered, nested representation of knowledge in deep neural networks (DNNs) and enables the extraction of lower footprint submodels without the need of retraining. We further show that for linear maps our Ordered Dropout is equivalent to SVD. We employ this technique, along with a self-distillation methodology, in the realm of FL in a framework called FjORD. FjORD alleviates the problem of client system heterogeneity by tailoring the model width to the client's capabilities. Extensive evaluation on both CNNs and RNNs across diverse modalities shows that FjORD consistently leads to significant performance gains over state-of-the-art baselines, while maintaining its nested structure.

cs.LG↗

arXiv2022

DynO: Dynamic Onloading of Deep Neural Networks from Cloud to Device

Recently, there has been an explosive growth of mobile and embedded applications using convolutional neural networks(CNNs). To alleviate their excessive computational demands, developers have traditionally resorted to cloud offloading, inducing high infrastructure costs and a strong dependence on networking conditions. On the other end, the emergence of powerful SoCs is gradually enabling on-device execution. Nonetheless, low- and mid-tier platforms still struggle to run state-of-the-art CNNs sufficiently. In this paper, we present DynO, a distributed inference framework that combines the best of both worlds to address several challenges, such as device heterogeneity, varying bandwidth and multi-objective requirements. Key components that enable this are its novel CNN-specific data packing method, which exploits the variability of precision needs in different parts of the CNN when onloading computation, and its novel scheduler that jointly tunes the partition point and transferred data precision at run time to adapt inference to its execution environment. Quantitative evaluation shows that DynO outperforms the current state-of-the-art, improving throughput by over an order of magnitude over device-only execution and up to 7.9x over competing CNN offloading systems, with up to 60x less data transferred.

cs.DC↗

arXiv2021

Smart at what cost? Characterising Mobile Deep Neural Networks in the wild

With smartphones' omnipresence in people's pockets, Machine Learning (ML) on mobile is gaining traction as devices become more powerful. With applications ranging from visual filters to voice assistants, intelligence on mobile comes in many forms and facets. However, Deep Neural Network (DNN) inference remains a compute intensive workload, with devices struggling to support intelligence at the cost of responsiveness.On the one hand, there is significant research on reducing model runtime requirements and supporting deployment on embedded devices. On the other hand, the strive to maximise the accuracy of a task is supported by deeper and wider neural networks, making mobile deployment of state-of-the-art DNNs a moving target. In this paper, we perform the first holistic study of DNN usage in the wild in an attempt to track deployed models and match how these run on widely deployed devices. To this end, we analyse over 16k of the most popular apps in the Google Play Store to characterise their DNN usage and performance across devices of different capabilities, both across tiers and generations. Simultaneously, we measure the models' energy footprint, as a core cost dimension of any mobile deployment. To streamline the process, we have developed gaugeNN, a tool that automates the deployment, measurement and analysis of DNNs on devices, with support for different frameworks and platforms. Results from our experience study paint the landscape of deep learning deployments on smartphones and indicate their popularity across app developers. Furthermore, our study shows the gap between bespoke techniques and real-world deployments and the need for optimised deployment of deep learning models in a highly dynamic and heterogeneous ecosystem.

cs.LG↗

arXiv2021

Disturbed YouTube for Kids: Characterizing and Detecting Inappropriate Videos Targeting Young Children

A large number of the most-subscribed YouTube channels target children of a very young age. Hundreds of toddler-oriented channels on YouTube feature inoffensive, well-produced, and educational videos. Unfortunately, inappropriate content that targets this demographic is also common. YouTube's algorithmic recommendation system regrettably suggests inappropriate content because some of it mimics or is derived from otherwise appropriate content. Considering the risk for early childhood development, and an increasing trend in toddler's consumption of YouTube media, this is a worrisome problem. In this work, we build a classifier able to discern inappropriate content that targets toddlers on YouTube with 84.3% accuracy, and leverage it to perform a first-of-its-kind, large-scale, quantitative characterization that reveals some of the risks of YouTube media consumption by young children. Our analysis reveals that YouTube is still plagued by such disturbing videos and its currently deployed counter-measures are ineffective in terms of detecting them in a timely manner. Alarmingly, using our classifier we show that young children are not only able, but likely to encounter disturbing videos when they randomly browse the platform starting from benign videos.

cs.SI↗

arXiv2021

How to Reach Real-Time AI on Consumer Devices? Solutions for Programmable and Custom Architectures

The unprecedented performance of deep neural networks (DNNs) has led to large strides in various Artificial Intelligence (AI) inference tasks, such as object and speech recognition. Nevertheless, deploying such AI models across commodity devices faces significant challenges: large computational cost, multiple performance objectives, hardware heterogeneity and a common need for high accuracy, together pose critical problems to the deployment of DNNs across the various embedded and mobile devices in the wild. As such, we have yet to witness the mainstream usage of state-of-the-art deep learning algorithms across consumer devices. In this paper, we provide preliminary answers to this potentially game-changing question by presenting an array of design techniques for efficient AI systems. We start by examining the major roadblocks when targeting both programmable processors and custom accelerators. Then, we present diverse methods for achieving real-time performance following a cross-stack approach. These span model-, system- and hardware-level techniques, and their combination. Our findings provide illustrative examples of AI systems that do not overburden mobile hardware, while also indicating how they can improve inference accuracy. Moreover, we showcase how custom ASIC- and FPGA-based accelerators can be an enabling factor for next-generation AI applications, such as multi-DNN systems. Collectively, these results highlight the critical need for further exploration as to how the various cross-stack solutions can be best combined in order to bring the latest advances in deep learning close to users, in a robust and efficient manner.

cs.LG↗

arXiv2021

It's always personal: Using Early Exits for Efficient On-Device CNN Personalisation

On-device machine learning is becoming a reality thanks to the availability of powerful hardware and model compression techniques. Typically, these models are pretrained on large GPU clusters and have enough parameters to generalise across a wide variety of inputs. In this work, we observe that a much smaller, personalised model can be employed to fit a specific scenario, resulting in both higher accuracy and faster execution. Nevertheless, on-device training is extremely challenging, imposing excessive computational and memory requirements even for flagship smartphones. At the same time, on-device data availability might be limited and samples are most frequently unlabelled. To this end, we introduce PersEPhonEE, a framework that attaches early exits on the model and personalises them on-device. These allow the model to progressively bypass a larger part of the computation as more personalised data become available. Moreover, we introduce an efficient on-device algorithm that trains the early exits in a semi-supervised manner at a fraction of the whole network's personalisation time. Results show that PersEPhonEE boosts accuracy by up to 15.9% while dropping the training cost by up to 2.2x and inference latency by 2.2-3.2x on average for the same accuracy, depending on the availability of labels on-device.

cs.LG↗

arXiv2020

SPINN: Synergistic Progressive Inference of Neural Networks over Device and Cloud

Despite the soaring use of convolutional neural networks (CNNs) in mobile applications, uniformly sustaining high-performance inference on mobile has been elusive due to the excessive computational demands of modern CNNs and the increasing diversity of deployed devices. A popular alternative comprises offloading CNN processing to powerful cloud-based servers. Nevertheless, by relying on the cloud to produce outputs, emerging mission-critical and high-mobility applications, such as drone obstacle avoidance or interactive applications, can suffer from the dynamic connectivity conditions and the uncertain availability of the cloud. In this paper, we propose SPINN, a distributed inference system that employs synergistic device-cloud computation together with a progressive inference method to deliver fast and robust CNN inference across diverse settings. The proposed system introduces a novel scheduler that co-optimises the early-exit policy and the CNN splitting at run time, in order to adapt to dynamic conditions and meet user-defined service-level requirements. Quantitative evaluation illustrates that SPINN outperforms its state-of-the-art collaborative inference counterparts by up to 2x in achieved throughput under varying network conditions, reduces the server cost by up to 6.8x and improves accuracy by 20.7% under latency constraints, while providing robust operation under uncertain connectivity conditions and significant energy savings compared to cloud-centric execution.

cs.LG↗

arXiv2020

A Retrospective Analysis of User Exposure to (Illicit) Cryptocurrency Mining on the Web

In late 2017, a sudden proliferation of malicious JavaScript was reported on the Web: browser-based mining exploited the CPU time of website visitors to mine the cryptocurrency Monero. Several studies measured the deployment of such code and developed defenses. However, previous work did not establish how many users were really exposed to the identified mining sites and whether there was a real risk given common user browsing behavior. In this paper, we present a retroactive analysis to close this research gap. We pool large-scale, longitudinal data from several vantage points, gathered during the prime time of illicit cryptomining, to measure the impact on web users. We leverage data from passive traffic monitoring of university networks and a large European ISP, with suspected mining sites identified in previous active scans. We corroborate our results with data from a browser extension with a large user base that tracks site visits. We also monitor open HTTP proxies and the Tor network for malicious injection of code. We find that the risk for most Web users was always very low, much lower than what deployment scans suggested. Any exposure period was also very brief. However, we also identify a previously unknown and exploited attack vector on mobile devices.

cs.CR↗

arXiv2020

DarkneTZ: Towards Model Privacy at the Edge using Trusted Execution Environments

We present DarkneTZ, a framework that uses an edge device's Trusted Execution Environment (TEE) in conjunction with model partitioning to limit the attack surface against Deep Neural Networks (DNNs). Increasingly, edge devices (smartphones and consumer IoT devices) are equipped with pre-trained DNNs for a variety of applications. This trend comes with privacy risks as models can leak information about their training data through effective membership inference attacks (MIAs). We evaluate the performance of DarkneTZ, including CPU execution time, memory usage, and accurate power consumption, using two small and six large image classification models. Due to the limited memory of the edge device's TEE, we partition model layers into more sensitive layers (to be executed inside the device TEE), and a set of layers to be executed in the untrusted part of the operating system. Our results show that even if a single layer is hidden, we can provide reliable model privacy and defend against state of the art MIAs, with only 3% performance overhead. When fully utilizing the TEE, DarkneTZ provides model protections with up to 10% overhead.

cs.LG↗

arXiv2019

"You Know What to Do": Proactive Detection of YouTube Videos Targeted by Coordinated Hate Attacks

Video sharing platforms like YouTube are increasingly targeted by aggression and hate attacks. Prior work has shown how these attacks often take place as a result of "raids," i.e., organized efforts by ad-hoc mobs coordinating from third-party communities. Despite the increasing relevance of this phenomenon, however, online services often lack effective countermeasures to mitigate it. Unlike well-studied problems like spam and phishing, coordinated aggressive behavior both targets and is perpetrated by humans, making defense mechanisms that look for automated activity unsuitable. Therefore, the de-facto solution is to reactively rely on user reports and human moderation. In this paper, we propose an automated solution to identify YouTube videos that are likely to be targeted by coordinated harassers from fringe communities like 4chan. First, we characterize and model YouTube videos along several axes (metadata, audio transcripts, thumbnails) based on a ground truth dataset of videos that were targeted by raids. Then, we use an ensemble of classifiers to determine the likelihood that a video will be raided with very good results (AUC up to 94%). Overall, our work provides an important first step towards deploying proactive systems to detect and mitigate coordinated hate attacks on platforms like YouTube.

cs.CY↗

arXiv2019

Detecting Cyberbullying and Cyberaggression in Social Media

Cyberbullying and cyberaggression are increasingly worrisome phenomena affecting people across all demographics. More than half of young social media users worldwide have been exposed to such prolonged and/or coordinated digital harassment. Victims can experience a wide range of emotions, with negative consequences such as embarrassment, depression, isolation from other community members, which embed the risk to lead to even more critical consequences, such as suicide attempts. In this work, we take the first concrete steps to understand the characteristics of abusive behavior in Twitter, one of today's largest social media platforms. We analyze 1.2 million users and 2.1 million tweets, comparing users participating in discussions around seemingly normal topics like the NBA, to those more likely to be hate-related, such as the Gamergate controversy, or the gender pay inequality at the BBC station. We also explore specific manifestations of abusive behavior, i.e., cyberbullying and cyberaggression, in one of the hate-related communities (Gamergate). We present a robust methodology to distinguish bullies and aggressors from normal Twitter users by considering text, user, and network-based attributes. Using various state-of-the-art machine learning algorithms, we classify these accounts with over 90% accuracy and AUC. Finally, we discuss the current status of Twitter user accounts marked as abusive by our methodology, and study the performance of potential mechanisms that can be used by Twitter to suspend users in the future.

cs.SI↗

arXiv2019

EmBench: Quantifying Performance Variations of Deep Neural Networks across Modern Commodity Devices

In recent years, advances in deep learning have resulted in unprecedented leaps in diverse tasks spanning from speech and object recognition to context awareness and health monitoring. As a result, an increasing number of AI-enabled applications are being developed targeting ubiquitous and mobile devices. While deep neural networks (DNNs) are getting bigger and more complex, they also impose a heavy computational and energy burden on the host devices, which has led to the integration of various specialized processors in commodity devices. Given the broad range of competing DNN architectures and the heterogeneity of the target hardware, there is an emerging need to understand the compatibility between DNN-platform pairs and the expected performance benefits on each platform. This work attempts to demystify this landscape by systematically evaluating a collection of state-of-the-art DNNs on a wide variety of commodity devices. In this respect, we identify potential bottlenecks in each architecture and provide important guidelines that can assist the community in the co-design of more efficient DNNs and accelerators.

cs.LG↗