Wainakh, Aidmar (2022)
On Privacy-Enhanced Distributed Analytics in Online Social Networks.
Technische Universität Darmstadt
doi: 10.26083/tuprints-00021034
Ph.D. Thesis, Primary publication, Publisher's Version
Text
2022-02-07_Wainakh_Aidmar.pdf Copyright Information: CC BY-SA 4.0 International - Creative Commons, Attribution ShareAlike. Download (8MB) |
Item Type: | Ph.D. Thesis | ||||
---|---|---|---|---|---|
Type of entry: | Primary publication | ||||
Title: | On Privacy-Enhanced Distributed Analytics in Online Social Networks | ||||
Language: | English | ||||
Referees: | Mühlhäuser, Prof. Dr. Max ; Fischer, Prof. Dr. Mathias | ||||
Date: | 2022 | ||||
Place of Publication: | Darmstadt | ||||
Collation: | xviii, 184 Seiten | ||||
Date of oral examination: | 21 March 2022 | ||||
DOI: | 10.26083/tuprints-00021034 | ||||
Abstract: | More than half of the world's population benefits from online social network (OSN) services. A considerable part of these services is mainly based on applying analytics on user data to infer their preferences and enrich their experience accordingly. At the same time, user data is monetized by service providers to run their business models. Therefore, providers tend to extensively collect (personal) data about users. However, this data is oftentimes used for various purposes without informed consent of the users. Providers share this data in different forms with third parties (e.g., data brokers). Moreover, user sensitive data was repeatedly a subject of unauthorized access by malicious parties. These issues have demonstrated the insufficient commitment of providers to user privacy, and consequently, raised users' concerns. Despite the emergence of privacy regulations (e.g., GDPR and CCPA), recent studies showed that user personal data collection and sharing sensitive data are still continuously increasing. A number of privacy-friendly OSNs have been proposed to enhance user privacy by reducing the need for central service providers. However, this improvement in privacy protection usually comes at the cost of losing social connectivity and many analytics-based services of the wide-spread OSNs. This dissertation addresses this issue by first proposing an approach to privacy-friendly OSNs that maintains established social connections. Second, approaches that allow users to collaboratively apply distributed analytics while preserving their privacy are presented. Finally, the dissertation contributes to better assessment and mitigation of the risks associated with distributed analytics. These three research directions are treated through the following six contributions. Conceptualizing Hybrid Online Social Networks: We conceptualize a hybrid approach to privacy-friendly OSNs, HOSN. This approach combines the benefits of using COSNs and DOSN. Users can maintain their social experience in their preferred COSN while being provided with additional means to enhance their privacy. Users can seamlessly post public content or private content that is accessible only by authorized users (friends) beyond the reach of the service providers. Improving the Trustworthiness of HOSNs: We conceptualize software features to address users' privacy concerns in OSNs. We prototype these features in our HOSN}approach and evaluate their impact on the privacy concerns and the trustworthiness of the approach. Also, we analyze the relationships between four important aspects that influence users' behavior in OSNs: privacy concerns, trust beliefs, risk beliefs, and the willingness to use. Privacy-Enhanced Association Rule Mining: We present an approach to enable users to apply efficiently privacy-enhanced association rule mining on distributed data. This approach can be employed in DOSN and HOSN to generate recommendations. We leverage a privacy-enhanced distributed graph sampling method to reduce the data required for the mining and lower the communication and computational overhead. Then, we apply a distributed frequent itemset mining algorithm in a privacy-friendly manner. Privacy Enhancements on Federated Learning (FL): We identify several privacy-related issues in the emerging distributed machine learning technique, FL. These issues are mainly due to the centralized nature of this technique. We discuss tackling these issues by applying FL in a hierarchical architecture. The benefits of this approach include a reduction in the centralization of control and the ability to place defense and verification methods more flexibly and efficiently within the hierarchy. Systematic Analysis of Threats in Federated Learning: We conduct a critical study of the existing attacks in FL to better understand the actual risk of these attacks under real-world scenarios. First, we structure the literature in this field and show the research foci and gaps. Then, we highlight a number of issues in (1) the assumptions commonly made by researchers and (2) the evaluation practices. Finally, we discuss the implications of these issues on the applicability of the proposed attacks and recommend several remedies. Label Leakage from Gradients: We identify a risk of information leakage when sharing gradients in FL. We demonstrate the severity of this risk by proposing a novel attack that extracts the user annotations that describe the data (i.e., ground-truth labels) from gradients. We show the high effectiveness of the attack under different settings such as different datasets and model architectures. We also test several defense mechanisms to mitigate this attack and conclude the effective ones. |
||||
Alternative Abstract: |
|
||||
Status: | Publisher's Version | ||||
URN: | urn:nbn:de:tuda-tuprints-210340 | ||||
Classification DDC: | 000 Generalities, computers, information > 004 Computer science | ||||
Divisions: | 20 Department of Computer Science > Telecooperation | ||||
TU-Projects: | DFG|GRK 2050|TP_Reuter_GRK_2050_a | ||||
Date Deposited: | 25 Mar 2022 13:11 | ||||
Last Modified: | 04 Aug 2022 09:10 | ||||
URI: | https://tuprints.ulb.tu-darmstadt.de/id/eprint/21034 | ||||
PPN: | 494261781 | ||||
Export: |
View Item |