TL;DR
This paper introduces TAPESTRY, a novel blockchain-based service that leverages users' "digital personhood" (longitudinal interaction data) to provide verifiable trust evidence for online identities. It aims to empower non-expert users to make better trust decisions by offering granular, privacy-preserving sharing of this evidence, facilitated by deep learning and a hybrid on/off-chain architecture.
What Problem Does This Paper Address?
The internet is rife with online fraud and scams, often perpetrated by ephemeral false identities. This makes it incredibly difficult for everyday users to determine who to trust online, leading to risks of over-sharing personal data and falling victim to malicious actors. Traditional online security relies on simple identifiers like pseudonyms or email addresses, which lack inherent trustworthiness. The authors propose that a person's "digital personhood" – the rich tapestry of their lifelong digital interactions (social media posts, likes, etc.) – can serve as a more robust basis for evidencing identity provenance and trustworthiness. However, effectively harnessing this data for trust decisions presents significant challenges in signal processing, privacy, security, and infrastructure, especially for users with low digital literacy.
Methods
TAPESTRY employs a multi-faceted approach:
- Digital Personhood (DP) Extraction: A deep neural network (DNN), specifically a triplet LSTM architecture, is used to process user-generated content (like Twitter posts). This DNN learns semantic embeddings of words and then models temporal coherence of activities over time to create "trust evidence" (TE) vectors. This process aims to distill complex interactions into a compact, privacy-preserving representation.
- Hybrid On- and Off-Chain Architecture: Encrypted trust evidence is stored in an off-chain "data lake" for privacy and GDPR compliance (allowing deletion). Hashes of this encrypted evidence, along with metadata, are stored immutably on a proof-of-work (PoW) blockchain to ensure integrity and provenance. This hybrid approach balances privacy with verifiable trust.
- Granular, Privacy-Preserving Sharing: Users can selectively share trust evidence with others. A symmetric key sharing scheme allows for granular control over what evidence is shared, to whom, and for what time periods.
- Data Visualisation: To make complex trust evidence comprehensible to non-experts, TAPESTRY uses novel visualisation techniques (e.g., "pie" and "slash" designs). These visualisations aim to convey the completeness and coherence of a user's digital history, helping verifiers make informed trust decisions without needing to understand the underlying technical details.
Datasets
The paper mentions experiments using a public dataset of Twitter posts ("tweets") gathered by Li et al. [45], initially comprising 50 million tweets from 140 thousand users. For their experiments, they filtered this dataset to include users with at least 800 tweets, resulting in a dataset of 8000 users for anomaly detection experiments.
Key Insights and Ideas
- Digital Personhood as Trust Evidence: The core idea that a person's accumulated digital interactions can serve as a verifiable basis for trust is a powerful shift from traditional identity verification.
- Balancing Privacy and Verifiability: The hybrid on/off-chain architecture is crucial for addressing the tension between the need for immutable provenance (blockchain) and the right to data deletion/privacy (off-chain data lake).
- Empowering Non-Experts: The emphasis on intuitive visualisations and user control over data sharing is key to making advanced security concepts accessible to a broader audience.
- Beyond Authentication: TAPESTRY focuses on the provenance and trustworthiness of an identity, rather than just authenticating that a user is who they claim to be.
Key Results
- Anomaly Detection: The proposed DNN approach outperformed baseline methods (OCSVM, Isolation Forest, LOF) in detecting anomalous behaviour in social media feeds, achieving a precision of 0.95 and recall of 0.91-0.95, and was invariant to the length of simulated hijacking.
- User Classification: The LSTM model achieved higher user classification accuracy compared to RNN and GRU models, especially as the number of users increased.
- Crowdfunding Experiment: Users with access to TAPESTRY visualisations made investment decisions in a mock crowdfunding scenario approximately twice as quickly as a control group, while achieving similar accuracy in distinguishing real from fake profiles.
- Visualisation Comprehensibility: An evaluation on Amazon Mechanical Turk showed that users could identify unusual activity patterns in social media feeds using the "slash" visualisation with high accuracy (~94%).
Limitations and Future Work
- Scalability: The authors acknowledge that at-scale deployment will require managing multiple data lakes and further characterization of the TAPESTRY blockchain's performance.
- Visualisation Interpretation: While designed for non-experts, the visualisations still necessitate some initial explanation and guidance. The paper notes that interpreting colour and tone can be subjective.
- Non-Semantic Nature of TE: The trust evidence (TE) is derived from semantic content but is itself non-semantic, which presents challenges in interpretation and understanding.
- Future Work: The paper suggests exploring at-scale deployment, further performance characterization, and potentially extending the system to support more complex trust scenarios.
Summary updated