arXiv subjects

Fraser Kennedy

Publications and source records attributed to Fraser Kennedy.

6 recordsLinked to original sources

arXiv2026

Exact Certification of Data-Poisoning Attacks Using Mixed-Integer Programming

This work introduces a verification framework that provides both sound and complete guarantees for data poisoning attacks during neural network training. We formulate adversarial data manipulation, model training, and test-time evaluation in a single mixed-integer quadratic programming (MIQCP) problem. Finding the global optimum of the proposed formulation provably yields worst-case poisoning attacks, while simultaneously bounding the effectiveness of all possible attacks on the given training pipeline. Our framework encodes both the gradient-based training dynamics and model evaluation at test time, enabling the first exact certification of training-time robustness. Experimental evaluation on small models confirms that our approach delivers a complete characterization of robustness against data poisoning.

cs.LG↗

arXiv2025

Detection of AI Generated Images Using Combined Uncertainty Measures and Particle Swarm Optimised Rejection Mechanism

As AI-generated images become increasingly photorealistic, distinguishing them from natural images poses a growing challenge. This paper presents a robust detection framework that leverages multiple uncertainty measures to decide whether to trust or reject a model's predictions. We focus on three complementary techniques: Fisher Information, which captures the sensitivity of model parameters to input variations; entropy-based uncertainty from Monte Carlo Dropout, which reflects predictive variability; and predictive variance from a Deep Kernel Learning framework using a Gaussian Process classifier. To integrate these diverse uncertainty signals, Particle Swarm Optimisation is used to learn optimal weightings and determine an adaptive rejection threshold. The model is trained on Stable Diffusion-generated images and evaluated on GLIDE, VQDM, Midjourney, BigGAN, and StyleGAN3, each introducing significant distribution shifts. While standard metrics such as prediction probability and Fisher-based measures perform well in distribution, their effectiveness degrades under shift. In contrast, the Combined Uncertainty measure consistently achieves an incorrect rejection rate of approximately 70 percent on unseen generators, successfully filtering most misclassified AI samples. Although the system occasionally rejects correct predictions from newer generators, this conservative behaviour is acceptable, as rejected samples can support retraining. The framework maintains high acceptance of accurate predictions for natural images and in-domain AI data. Under adversarial attacks using FGSM and PGD, the Combined Uncertainty method rejects around 61 percent of successful attacks, while GP-based uncertainty alone achieves up to 80 percent. Overall, the results demonstrate that multi-source uncertainty fusion provides a resilient and adaptive solution for AI-generated image detection.

cs.CV↗

arXiv2024

Sensitivity of Bayesian 21 cm power spectrum estimation to foreground model errors

Power spectrum estimators are an important tool in efforts to detect the 21 cm brightness temperature fluctuations from neutral hydrogen at early times. An initial detection will likely be statistical in nature, meaning that it will not be possible to make a coherent map of the brightness temperature fluctuations; instead, only their variance will be measured against a background of noise and residual systematic effects. Optimal Quadratic Estimator (OQE)-based methods often apply an inverse covariance weighting to the data. However, inaccurate covariance modelling can lead to reduced sensitivity and, in some cases, severe signal loss. We recently proposed a Bayesian method to jointly estimate the 21 cm fluctuations, their power spectrum, and foreground emission. Instead of requiring a fixed a priori estimate of the covariance, we estimate the covariance as part of the inference. Choices of parametrization, particularly of the foregrounds, are subject to model errors and could lead to biases and other ill effects if not properly controlled. In this paper, we investigate the effects of inaccurate foreground models on 21 cm power spectrum recovery. Using simulated visibilities, we find that, even in the most extreme scenarios tested, our approach is capable of recovering 21 cm delay power spectrum estimates consistent with a known input signal for delays $\gtrsim300$ ns ($\sim$88\% of the available Fourier modes). This is true even when using foreground models derived from modified foreground catalogs containing spatial and spectral perturbations at the quoted level of uncertainty on our foreground catalogs.

astro-ph.CO↗

arXiv2023

Statistical recovery of 21cm visibilities and their power spectra with Gaussian constrained realisations and Gibbs sampling

Radio interferometers designed to probe the 21cm signal from Cosmic Dawn and the Epoch of Reionisation must contend with systematic effects that make it difficult to achieve sufficient dynamic range to separate the 21cm signal from foreground emission and other effects. For instance, the instrument's chromatic response modulates the otherwise spectrally smooth foregrounds, making them difficult to model, while a significant fraction of the data must be excised due to the presence of radio frequency interference (RFI), leaving gaps in the data. Errors in modelling the (modulated and gappy) foregrounds can easily generate spurious contamination of what should otherwise be 21cm signal-dominated modes. Various approaches have been developed to mitigate these issues by (e.g.) using non-parametric reconstruction of the foregrounds, in-painting the gaps, and weighting the data to reduce the level of contamination. We present a Bayesian statistical method that combines these approaches, using the coupled techniques of Gaussian constrained realisations (GCR) and Gibbs sampling. This provides a way of drawing samples from the joint posterior distribution of the 21cm signal modes and their power spectrum in the presence of gappy data and an uncertain foreground model in a computationally scalable manner. The data are weighted by an inverse covariance matrix that is estimated as part of the inference, along with a foreground model that can then be marginalised over. We demonstrate the application of this technique on a simulated HERA-like delay spectrum analysis, comparing three different approaches for accounting for the foreground components.

astro-ph.CO↗

arXiv2022

Bayesian jackknife tests with a small number of subsets: Application to HERA 21cm power spectrum upper limits

We present a Bayesian jackknife test for assessing the probability that a data set contains biased subsets, and, if so, which of the subsets are likely to be biased. The test can be used to assess the presence and likely source of statistical tension between different measurements of the same quantities in an automated manner. Under certain broadly applicable assumptions, the test is analytically tractable. We also provide an open source code, CHIBORG, that performs both analytic and numerical computations of the test on general Gaussian-distributed data. After exploring the information theoretical aspects of the test and its performance with an array of simulations, we apply it to data from the Hydrogen Epoch of Reionization Array (HERA) to assess whether different sub-seasons of observing can justifiably be combined to produce a deeper 21cm power spectrum upper limit. We find that, with a handful of exceptions, the HERA data in question are statistically consistent and this decision is justified. We conclude by pointing out the wide applicability of this test, including to CMB experiments and the $H_0$ tension.

astro-ph.IM↗

arXiv2021

Statistical recovery of the BAO scale from multipoles of the beam-convolved 21cm correlation function

Despite being designed as an interferometer, the MeerKAT radio array (an SKA pathfinder) can also be used in autocorrelation (`single-dish') mode, where each dish scans the sky independently. Operating in this mode allows extremely high survey speeds to be achieved, albeit at significantly lower angular resolution. We investigate the recovery of the baryon acoustic oscillation (BAO) scale from multipoles of the redshift-space correlation function as measured by a low angular resolution 21cm intensity mapping survey of this kind. Our approach is to construct an analytic model of the multipoles of the correlation function and their covariance matrix that includes foreground contamination and beam resolution effects, which we then use to generate an ensemble of mock data vectors from which we attempt to recover the BAO scale. In line with previous studies, we find that recovery of the transverse BAO scale $α_{\perp}$ is hampered by the strong smoothing effect of the instrumental beam with increasing redshift, while the radial scale $α_\parallel$ is much more robust. The multipole formalism naturally incorporates transverse information when it is available however, and so there is no need to perform a radial-only analysis. In particular, the quadrupole of the correlation function preserves a distinctive BAO `bump' feature even for large smoothing scales. We also investigate the robustness of BAO scale recovery to beam model accuracy, severity of the foreground removal cuts, and accuracy of the covariance matrix model, finding in all cases that the radial BAO scale can be recovered in an accurate, unbiased manner.

astro-ph.CO↗