We present COAST (consecutive multi-omics alignment of spatial tissues), a method to reliably physically align consecutive tissue sections to produce a unified multi-modal molecular dataset suitable for downstream applications. COAST relies exclusively on the images associated with spatial data, eliminating the need for common molecular features or prior annotations. We demonstrate the effectiveness of COAST using spatial transcriptomics slides from different technologies, tissues, and resolutions, in which it achieves performance comparable to established uni-modal alignment tools. Applying COAST to spatial transcriptomics and metabolomics/lipidomics tissue sections from a mouse model of ischemia-reperfusion injury allowed the investigation of lipid/metabolite features of transcriptionally defined cell types. Overall, COAST offers a streamlined and integrative solution for multi-modal spatial data alignment. ... Read more

Given the significant amount of time that people spend indoors, public spaces are reinventing themselves to deliver immersive experiences. One approach is to provide interactive gaming, such as the spaces created in big international airports like Singapore. The problem is that these gaming areas rely on either (i) wearable devices, which can easily spread germs, particularly in highly transited areas, such as those in airports, or (ii) cameras, which are increasingly raising privacy concerns and are being forbidden in some public areas. To provide an immersive gaming experience that is device-free and privacy-aware, we propose a system based on mmWave radar. In particular, our work provides three contributions. (1) For the first time, we compare various tracking and human pose estimation (HPE) models in the SoA under a unified framework. This approach allows us to identify the best methods in terms of position accuracy, latency, and smoothness-which are critical gaming metrics. (2) We build an integrated system with tracking and HPE capabilities. Our system includes the design of four games with different levels of complexity and a dataset collected specifically to train models for gaming applications. (3) Our evaluation, which includes a subjective users' survey and an objective comparison with a Kinect console, shows that today's mmWave is suitable for games where only the user's location is required or when coarse-grained gestures are needed. There are two key areas where mmWave research needs to improve to match camera-based consoles: sensors need a higher sampling rate, since a high frame rate is critical for seamless gaming, and HPE models need to master fine-grained gestures, in particular when arms end up in front of the body. All the code and datasets will be made available as a stepping stone for future work. ... Read more

Federated Learning enables decentralized clients to train a shared model without transferring their raw data to a central server. However, the aggregation server cannot directly inspect local training data or verify whether submitted updates were produced honestly. This makes Federated Learning vulnerable to backdoor attacks in which malicious clients embed a trigger-dependent misclassification while preserving ordinary model performance.

This work builds directly on the Robust Learning Rate defence of Ozdayi et al., whose IID Fashion-MNIST experiments reported 93.4% validation accuracy and 100% backdoor accuracy under attacked Federated Averaging, compared with 92.9% validation accuracy and 0% backdoor accuracy when Robust Learning Rate was applied. Our reproduction closely matches these results: the undefended static attack achieved a final-50-round Attack Success Rate of 100.000% with 93.405% validation accuracy, whereas Robust Learning Rate with threshold θ = 4 reduced final-window attack success to 0.000% while preserving 92.311% validation accuracy.

These results establish a stronger adversarial problem: an adaptive attack must retain the effectiveness of static poisoning while modifying its temporal and directional properties so that its malicious contribution is less immediately incompatible with honest-client updates. We therefore introduce Cuckoo, a temporally adaptive backdoor framework for Horizontal Federated Learning. Cuckoo separates training into a low-intensity incubation phase and a later hatching phase. In the principal configuration, the effective poisoning fraction increases from ρinc = 0.05 before round th = 100 to ρhatch = 0.50 afterwards.

The framework is evaluated across five datasets, six neural-network architectures, multiple poisoning budgets, attacker populations, federation sizes, activation rounds, data distributions, and server-side defences. In selected Fashion-MNIST scenarios, Cuckoo reached final-window Attack Success Rates between 99.77% and 100% while maintaining approximately 93.3–93.5% clean validation accuracy. A 50-client setting reached 99.77% final-window attack success, and a poisoning fraction of only 0.05 reached 99.90%. Under non-IID data, attack success remained 100%, although validation accuracy decreased to 87.49%.

Clean-label defence experiments produced a different result: Krum, coordinate-wise median, and strengthened median aggregation reduced final-window attack success to 2.13%, 8.11%, and 3.92%, respectively, with corresponding validation accuracies of 88.84%, 93.01%, and 92.63%. The results therefore do not show a universal defence bypass. Instead, they demonstrate that temporal attack effectiveness is strongly conditional on the attack type, data distribution, architecture, attacker concentration, and defence configuration. The findings motivate history-aware defences that evaluate client behaviour across communication rounds rather than treating each update as an isolated event. ... Read more

Digital technology plays an increasingly important role in practices surrounding death, grief, and remembrance. In this context, generative Artificial Intelligence (AI) is fundamentally reshaping how digital legacy is understood. Traditional forms of digital legacy are shifting towards interactive representations of the deceased, thereby fostering a sense of continuity after death. The literature refers to this development as “AI Afterlife.” However, while these representations may offer support during grief, they may also raise concerns such as dependency, problematic attachment, profit motives, privacy violations, and misrepresentation. Given these concerns, user-centred research is needed to understand how this technology can support both the represented deceased and the bereaved. However, due to the rapid development of these representations, empirical research has not kept pace. This study therefore examined how people perceive AI-generated representations of deceased persons as digital legacies. It also investigated what role relational proximity plays in shaping these perceptions. Additionally, the study examined what people needed to challenge or adjust a representation when they perceived it as problematic, with contestability at its centre. ... Read more

In this work, we extend nanoTabPFN, a Prior-Data Fitted Network (PFN) for tabular classification to produce Shapley attributions alongside its predictive distribution. The output is a bucketed distribution over signed feature attribution values. The class prediction is derived as the additive sum of a baseline and per-feature contributions, ensuring the explanation is faithful to the prediction by construction. The attribution head is supervised by TreeSHAP values from a decision tree prior, connecting the explanation to the data generating process (DGP). The resulting attributions are interpreted as a posterior over Shapley values. The PFN’s distributional mechanism is transferred to feature attributions to yield calibrated explanation uncertainty. Across synthetic and real tabular benchmarks, the model recovers attributions with high fidelity (R2 ≈ 0.94-0.95 against exact Shapley on synthetic data), but at a small cost to accuracy. It produces calibrated attribution uncertainty, with negative log-likelihood (NLL) below the uninformative uniform baseline. ... Read more