Transfer learning of diffusion models to smaller target domains is challenging, as naively fine-tuning the model often results in poor generalization. Test-time guidance methods help mitigate this by offering controllable improvements in image fidelity through a trade-off with sample diversity. However, this benefit comes at a high computational cost, typically requiring dual forward passes during sampling. We propose the Domain-guided Fine-tuning (DogFit) method, an effective guidance mechanism for diffusion transfer learning that maintains controllability without incurring additional computational overhead. DogFit injects a domain-aware guidance offset into the training loss, effectively internalizing the guided behavior during the fine-tuning process. The domain-aware design is motivated by our observation that during fine-tuning, the unconditional source model offers a stronger marginal estimate than the target model. To support efficient controllable fidelity–diversity trade-offs at inference, we encode the guidance strength value as an additional model input through a lightweight conditioning mechanism. We further investigate the optimal placement and timing of the guidance offset during training and propose two simple scheduling strategies, i.e., late-start and cut-off, which improve generation quality and training stability. Experiments on DiT and SiT backbones across six diverse target domains show that DogFit can outperform prior guidance methods in transfer learning in terms of FID and FD DINOV2 while requiring up to 2x fewer sampling TFLOPS.

Multimodal learning has shown promise in medical imaging, combining complementary modalities like images and text. Vision-language models (VLMs) capture rich diagnostic cues but often require large paired datasets and promptor text-based inference. Their practicality is therefore limited due to annotation cost, privacy, and compute demands. Unpaired external text, like pathology reports, can still provide complementary diagnostic cues if semantically relevant content is retrievable per image. To address this, we introduce CLIP-IT, a novel framework that relies on rich unpaired text reports. Specifically, CLIP-IT uses a CLIP model pre-trained on histology image–text pairs from a separate dataset to retrieve the most relevant unpaired textual report for each image in the downstream unimodal dataset. These reports, sourced from the same disease domain and tissue type, form pseudo-pairs that reflect shared clinical semantics rather than exact alignment. Knowledge from these texts is distilled into the vision model during training, while LoRA-based adaptation mitigates the semantic gap between unaligned modalities. At inference, only the vision model is used, maintaining low overhead while still benefiting from multimodal training without requiring paired data in the downstream dataset. Experiments1show that CLIP-IT consistently improves classification accuracy over both unimodal and multimodal CLIP-based baselines in most cases, without requiring paired annotations per dataset or incurring additional inference-time complexity.

Casting complex inputs into tractable representations is a critical step across various fields. Diverse embedding models emerge from differences in architectures, loss functions, input modalities and datasets, each capturing unique aspects of the input. Multi-teacher distillation leverages this diversity to enrich representations but often remains tailored to specific tasks. We introduce a task-agnostic framework based on a ``majority vote" objective function. We demonstrate that this function is bounded by the mutual information between the student and the teachers' embeddings, leading to a task-agnostic distillation loss that eliminates dependence on task-specific labels or prior knowledge. Comprehensive evaluations across text, vision models, and molecular modeling show that our method effectively leverages teacher diversity, resulting in representations enabling better performance for a wide range of downstream tasks such as classification, clustering, or regression. Additionally, we train and release state-of-the-art embedding models, enhancing downstream performance in various modalities.

Generative Adversarial Networks (GANs) have been widely used for generating synthetic data for cases where there is a limited size real-world data set or when data holders are unwilling to share their data samples. Recent works showed that GANs, due to overfitting and memorization, might leak information regarding their training data samples. This makes GANs vulnerable to Membership Inference Attacks (MIAs). Several defense strategies have been proposed in the literature to mitigate this privacy issue. Unfortunately, defense strategies based on differential privacy are proven to reduce extensively the quality of the synthetic data points. On the other hand, more recent frameworks such as PrivGAN and PAR-GAN are not suitable for small-size training data sets. In the present work, the overfitting in GANs is studied in terms of the discriminator, and a more general measure of overfitting based on the Bhattacharyya coefficient is defined. Then, inspired by Fano’s inequality, our first defense mechanism against MIAs is proposed. This framework, which requires only a simple modification in the loss function of GANs, is referred to as the maximum entropy GAN or MEGAN and significantly improves the robustness of GANs to MIAs. As a second defense strategy, a more heuristic model based on minimizing the information leaked from the generated samples about the training data points is presented. This approach is referred to as mutual information minimization GAN (MIMGAN) and uses a variational representation of the mutual information to minimize the information that a synthetic sample might leak about the whole training data set. Applying the proposed frameworks to some commonly used data sets against state-of-the-art MIAs reveals that the proposed methods can reduce the accuracy of the adversaries to the level of random guessing accuracy with a small reduction in the quality of the synthetic data samples.

Cardiotocography (CTG) plays a vital role in fetal well-being monitoring by tracking fetal heart rate (FHR) and uterine contractions (UC). However, CTG interpretation suffers from subjectivity, resulting in low agreement among observers and potentially unnecessary medical interventions. Existing AI-based diagnostic models struggle with generalization and are only effective on distinct CTG samples. This study introduces a novel approach employing deep semi-supervised learning for anomaly detection in CTG signals, marking the first attempt in this direction. A modified GANomaly model is proposed, trained on normal CTG data, and evaluated for abnormality detection. This model employs an encoder-decoder structure with a discriminator, minimizing reconstruction and latent space errors while learning the normal CTG signal distribution. Leveraging the CTU-UHB dataset, our model demonstrates superior performance compared to existing methods during inference.

This paper adopts Arimoto's$\alpha$-Mutual Information as a tunable privacy measure, in a privacy-preserving data release setting that aims to prevent disclosing private data to adversaries. By fine-tuning the privacy metric, we demonstrate that our approach yields superior models that effectively thwart attackers across various performance dimensions. We formulate a general distortion-based mechanism that manipulates the original data to offer privacy protection. The distortion metrics are determined according to the data structure of a specific experiment. We confront the problem expressed in the formulation by employing a general adversarial deep learning framework that consists of a releaser and an adversary, trained with opposite goals. This study conducts empirical experiments on images and time-series data to verify the functionality of$\alpha$-Mutual Information. We evaluate the privacy-utility trade-off of customized models and compare them to mutual information as the baseline measure. Finally, we analyze the consequence of an attacker's access to side information about private data and witness that adapting the privacy measure results in a more refined model than the state-of-the-art in terms of resiliency against side information.

A privacy-preserving adversarial network (PPAN) was recently proposed as an information-theoretical framework to address the issue of privacy in data sharing. The main idea of this model was to use mutual information as the privacy measure and adversarial training of two deep neural networks, one as the mechanism and another as the adversary. The performance of the PPAN model for the discrete synthetic data, MNIST handwritten digits, and continuous Gaussian data was evaluated and compared to the analytically optimal trade-off. In this study, we evaluate the PPAN model for continuous non-Gaussian data, i.e. smart meters data, where lower and upper bounds of the privacy-preserving problem are used. These bounds include the Kraskov (KSG) estimation of entropy and mutual information that is based on the k-th nearest neighbor. In addition to the synthetic data sets, a practical case for hiding the actual electricity consumption from smart meter readings is examined. The results show that for continuous non-Gaussian data, the PPAN model performs within the determined optimal ranges and close to the lower bound.

The rapid development and expansion of the Internet of Things (IoT) paradigm has drastically increased the collection and exchange of data between sensors and systems, a phenomenon that raises serious privacy concerns in some domains. In particular, smart meters (SMs) share fine-grained electricity consumption of households with utility providers that can potentially violate users’ privacy as sensitive information is leaked through the data. In order to enhance privacy, electricity consumers can exploit the availability of physical resources such as a rechargeable battery (RB) to shape their power demand as dictated by a privacy-cost management unit (PCMU). In this article, we present a novel method to learn the PCMU policy using deep reinforcement learning (DRL). We adopt the mutual information (MI) between the user’s demand load and the masked load seen by the power grid as a reliable and general privacy measure. Unlike previous studies, we model the whole temporal correlation in the data to learn the MI in its general form and use a neural network to estimate the MI-based reward signal to guide the PCMU learning process. This approach is combined with a model-free DRL algorithm known as the deep double${Q}$-learning (DDQL) method. The performance of the complete DDQL-MI algorithm is assessed empirically using an actual SMs data set and compared with simpler privacy measures. Our results show significant improvements over state-of-the-art privacy-aware demand shaping methods.