arXiv subjects

Godwin Badu-Marfo

Publications and source records attributed to Godwin Badu-Marfo.

12 recordsLinked to original sources

arXiv2026

The Interlocutor Effect: Why LLMs Leak More Personal Data to Agents Than Humans

Large Language Models (LLMs) alter their privacy behavior based on the perceived identity of their interlocutor. While safety mechanisms typically prevent LLMs from releasing Personally Identifiable Information (PII) to human users, these models tend to reveal more sensitive data when addressing another AI agent. We refer to this as the \textbf{Interlocutor Effect}. Through an ablation study, we find evidence that the technical nature of the recipient contributes to this effect, thereby diminishing the model's caution regarding privacy. To explore this further, we introduce the Attention Suppression Hypothesis, which posits that safety-aligned attention heads become inactive during interactions with agents. We assess this quantitatively by comparing human-directed and agent-directed prompts in 222 sensitive scenarios. Our findings, drawn from 3,464 interactions, indicate that portraying the recipient as an AI agent elevates PII leakage by up to 23 percentage points. Initial experiments on Llama-3.1-8B-Instruct corroborate this: deactivating one safety head induces leakage, whereas reactivating it reinstates privacy safeguards. We consider the implications for developing secure multi-agent systems.

cs.HC↗

arXiv2026

AgentLeak: A Benchmark for Internal-Channel Privacy Leakage in Multi-Agent LLM Systems

Multi-agent Large Language Model (LLM) systems create privacy risks that current output-only benchmarks cannot measure. When agents coordinate on tasks, sensitive data may pass through inter-agent messages, shared memory, and tool arguments, all pathways that final-output audits typically do not inspect. We introduce AgentLeak, a benchmark for evaluating internal-channel privacy leakage in multi-agent LLM systems. AgentLeak instruments seven privacy-relevant communication pathways and provides a large-scale empirical evaluation focused on final outputs, inter-agent messages, and shared memory. Across 1,000 scenarios spanning healthcare, finance, legal, and corporate domains, five production LLMs (GPT-4o, GPT-4o-mini, Claude 3.5 Sonnet, Mistral Large, and Llama 3.3 70B), and 4,979 validated execution traces, we find that multi-agent configurations reduce final-output leakage (C1: 27.2% vs 43.2% in single-agent mode) compared with single-agent baselines but introduce internal channels that raise total system exposure to 68.9% (aggregated across C1, C2, C5). Inter-agent messages (C2) leak at 68.8%, compared with 27.2% for final outputs (C1), meaning that output-only audits miss 41.7% of violations. Across all five models and four domains, the pattern C2 $\geq$ C1 holds consistently. These results suggest, within the evaluated coordinator-worker setting, that privacy risk in multi-agent systems is strongly shaped by architectural coordination channels rather than final-output behavior alone: it arises from internal channels that remain invisible to standard output-level defenses.

cs.AI↗

arXiv2025

Dynamic Campus Origin-Destination Mobility Prediction using Graph Convolutional Neural Network on WiFi Logs

We present an integrated graph-based neural networks architecture for predicting campus buildings occupancy and inter-buildings movement at dynamic temporal resolution that learns traffic flow patterns from Wi-Fi logs combined with the usage schedules within the buildings. The relative traffic flows are directly estimated from the WiFi data without assuming the occupant behaviour or preferences while maintaining individual privacy. We formulate the problem as a data-driven graph structure represented by a set of nodes (representing buildings), connected through a route of edges or links using a novel Graph Convolution plus LSTM Neural Network (GCLSTM) which has shown remarkable success in modelling complex patterns. We describe the formulation, model estimation, interpretability and examine the relative performance of our proposed model. We also present an illustrative architecture of the models and apply on real-world WiFi logs collected at the Toronto Metropolitan University campus. The results of the experiments show that the integrated GCLSTM models significantly outperform traditional pedestrian flow estimators like the Multi Layer Perceptron (MLP) and Linear Regression.

cs.LG↗

arXiv2024

Defense via Behavior Attestation against Attacks in Connected and Automated Vehicles based Federated Learning Systems

The recent application of Federated Learning algorithms in IOT and Wireless vehicular networks have given rise to newer cyber threats in the mobile environment which hitherto were not present in traditional fixed networks. These threats arise due to the intrinsic nature of wireless transmission medium and other inherent characteristics of mobile networks such as high-node mobility and rapidly changing topology. This paper investigates the robustness of Vehicular AttestedFL defense strategies against falsified information attacks by tracking the behavior. We show that the defense strategies are capable of detecting and eliminating malicious nodes in the wireless mobile setting of the future smart road networks.

eess.SY↗

arXiv2022

Robustness Analysis of Deep Learning Models for Population Synthesis

Deep generative models have become useful for synthetic data generation, particularly population synthesis. The models implicitly learn the probability distribution of a dataset and can draw samples from a distribution. Several models have been proposed, but their performance is only tested on a single cross-sectional sample. The implementation of population synthesis on single datasets is seen as a drawback that needs further studies to explore the robustness of the models on multiple datasets. While comparing with the real data can increase trust and interpretability of the models, techniques to evaluate deep generative models' robustness for population synthesis remain underexplored. In this study, we present bootstrap confidence interval for the deep generative models, an approach that computes efficient confidence intervals for mean errors predictions to evaluate the robustness of the models to multiple datasets. Specifically, we adopt the tabular-based Composite Travel Generative Adversarial Network (CTGAN) and Variational Autoencoder (VAE), to estimate the distribution of the population, by generating agents that have tabular data using several samples over time from the same study area. The models are implemented on multiple travel diaries of Montreal Origin- Destination Survey of 2008, 2013, and 2018 and compare the predictive performance under varying sample sizes from multiple surveys. Results show that the predictive errors of CTGAN have narrower confidence intervals indicating its robustness to multiple datasets of the varying sample sizes when compared to VAE. Again, the evaluation of model robustness against varying sample size shows a minimal decrease in model performance with decrease in sample size. This study directly supports agent-based modelling by enabling finer synthetic generation of populations in a reliable environment.

cs.LG↗

arXiv2022

eFedDNN: Ensemble based Federated Deep Neural Networks for Trajectory Mode Inference

As the most significant data source in smart mobility systems, GPS trajectories can help identify user travel mode. However, these GPS datasets may contain users' private information (e.g., home location), preventing many users from sharing their private information with a third party. Hence, identifying travel modes while protecting users' privacy is a significant issue. To address this challenge, we use federated learning (FL), a privacy-preserving machine learning technique that aims at collaboratively training a robust global model by accessing users' locally trained models but not their raw data. Specifically, we designed a novel ensemble-based Federated Deep Neural Network (eFedDNN). The ensemble method combines the outputs of the different models learned via FL by the users and shows an accuracy that surpasses comparable models reported in the literature. Extensive experimental studies on a real-world open-access dataset from Montreal demonstrate that the proposed inference model can achieve accurate identification of users' mode of travel without compromising privacy.

cs.LG↗

arXiv2021

On the Initial Behavior Monitoring Issues in Federated Learning

In Federated Learning (FL), a group of workers participate to build a global model under the coordination of one node, the chief. Regarding the cybersecurity of FL, some attacks aim at injecting the fabricated local model updates into the system. Some defenses are based on malicious worker detection and behavioral pattern analysis. In this context, without timely and dynamic monitoring methods, the chief cannot detect and remove the malicious or unreliable workers from the system. Our work emphasize the urgency to prepare the federated learning process for monitoring and eventually behavioral pattern analysis. We study the information inside the learning process in the early stages of training, propose a monitoring process and evaluate the monitoring period required. The aim is to analyse at what time is it appropriate to start the detection algorithm in order to remove the malicious or unreliable workers from the system and optimise the defense mechanism deployment. We tested our strategy on a behavioral pattern analysis defense applied to the FL process of different benchmark systems for text and image classification. Our results show that the monitoring process lowers false positives and false negatives and consequently increases system efficiency by enabling the distributed learning system to achieve better performance in the early stage of training.

cs.LG↗

arXiv2021

Cybersecurity Threats in Connected and Automated Vehicles based Federated Learning Systems

Federated learning (FL) is a machine learning technique that aims at training an algorithm across decentralized entities holding their local data private. Wireless mobile networks allow users to communicate with other fixed or mobile users. The road traffic network represents an infrastructure-based configuration of a wireless mobile network where the Connected and Automated Vehicles (CAV) represent the communicating entities. Applying FL in a wireless mobile network setting gives rise to a new threat in the mobile environment that is very different from the traditional fixed networks. The threat is due to the intrinsic characteristics of the wireless medium and is caused by the characteristics of the vehicular networks such as high node-mobility and rapidly changing topology. Most cyber defense techniques depend on highly reliable and connected networks. This paper explores falsified information attacks, which target the FL process that is ongoing at the RSU. We identified a number of attack strategies conducted by the malicious CAVs to disrupt the training of the global model in vehicular networks. We show that the attacks were able to increase the convergence time and decrease the accuracy the model. We demonstrate that our attacks bypass FL defense strategies in their primary form and highlight the need for novel poisoning resilience defense mechanisms in the wireless mobile setting of the future road networks.

cs.CR↗

arXiv2020

A Differentially Private Multi-Output Deep Generative Networks Approach For Activity Diary Synthesis

In this work, we develop a privacy-by-design generative model for synthesizing the activity diary of the travel population using state-of-art deep learning approaches. This proposed approach extends literature on population synthesis by contributing novel deep learning to the development and application of synthetic travel data while guaranteeing privacy protection for members of the sample population on which the synthetic populations are based. First, we show a complete de-generalization of activity diaries to simulate the socioeconomic features and longitudinal sequences of geographically and temporally explicit activities. Second, we introduce a differential privacy approach to control the level of resolution disclosing the uniqueness of survey participants. Finally, we experiment using the Generative Adversarial Networks (GANs). We evaluate the statistical distributions, pairwise correlations and measure the level of privacy guaranteed on simulated datasets for varying noise. The results of the model show successes in simulating activity diaries composed of multiple outputs including structured socio-economic features and sequential tour activities in a differentially private manner.

cs.LG↗

arXiv2020

Composite Travel Generative Adversarial Networks for Tabular and Sequential Population Synthesis

Agent-based transportation modelling has become the standard to simulate travel behaviour, mobility choices and activity preferences using disaggregate travel demand data for entire populations, data that are not typically readily available. Various methods have been proposed to synthesize population data for this purpose. We present a Composite Travel Generative Adversarial Network (CTGAN), a novel deep generative model to estimate the underlying joint distribution of a population, that is capable of reconstructing composite synthetic agents having tabular (e.g. age and sex) as well as sequential mobility data (e.g. trip trajectory and sequence). The CTGAN model is compared with other recently proposed methods such as the Variational Autoencoders (VAE) method, which has shown success in high dimensional tabular population synthesis. We evaluate the performance of the synthesized outputs based on distribution similarity, multi-variate correlations and spatio-temporal metrics. The results show the consistent and accurate generation of synthetic populations and their tabular and spatially sequential attributes, generated over varying spatial scales and dimensions.

cs.LG↗

arXiv2019

Perturbation Methods for Protection of Sensitive Location Data: Smartphone Travel Survey Case Study

Smartphone based travel data collection has become an important tool for the analysis of transportation systems. Interest in sharing travel survey data has gained popularity in recent years as "Open Data Initiatives" by governments seek to allow the public to use these data, and hopefully be able to contribute their findings and analysis to the public sphere. The public release of such precise information, particularly location data such as place of residence, opens the risk of privacy violation. At the same time, in order for such data to be useful, as much spatial resolution as possible is desirable for utility in transportation applications and travel demand modeling. This paper evaluates geographic random perturbation methods (i.e. Geo-indistinguishability and the Donut geomask) in protecting the privacy of respondents whose residential location may be published. We measure the performance of location privacy methods, preservation of utility and randomness in the distribution of perturbation distances with varying parameters. It is found that both methods produce distributions of spatial perturbations that conform closely to common probability distributions and as a result, that the original locations can be inferred with little information and a high degree of precision. It is also found that while Achieved K-estimate anonymity increases linearly with desired anonymity for the Donut geomask, Geo-Indistinguishability is highly dependent upon its privacy budget factor (epsilon) and is not very effective at assuring desired Achieved K-estimate anonymity.

cs.CR↗

arXiv2018

A Perspective on the Challenges and Opportunities for Privacy-Aware Big Transportation Data

In recent years, and especially since the development of the smartphone, enormous amounts of data relevant for transportation have become available. These data hold out the potential to redefine how transportation system (i.e. design, planning and operations) is done. While researchers in both academia and industry are making advances in using this data to transportation system ends (e.g. information inference from collected data), little attention has been paid to four larger scale challenges that will need to be overcome if the potential for Big Transportation Data is to be harnessed for transportation decision-making purposes. This paper aims to provide awareness of these large-scale challenges and provides insight into how we believe these challenges are likely to be met.

cs.CY↗