Download (v1.0, annotated N-triples)

We provide a sample dataset containing malicious triples that could be used to evaluate the resilience of Linked Data applications or to train spam filters.

The dataset is the polluted version of a fraction of the Billion Triple Challenge 2012 Dataset. More specifically, we chose the 1-hop expansion “Timbl crawl”, a crawl seeded with Tim Berners Lee’s foaf profile, and we applied the spam vectors described in the paper.

The resulting dataset contains approximately 16k triples (spam triples account for 4% of dataset size). The dataset includes samples of Content contamination vectors, Link poisoning vectors and Non-triple-based attacks (Malicious subclassing only).

• False labeling: rdfs:label properties associated to akt:Organisation entities
• Misattribution A malicious bibo:Quote has been associated to all foaf:Person of the dataset
• Void pollution: added deceiving dcterms:subject triples (about computer science publications)
• Schema pollution: all owl:Class and rdf:Property have been associated to malicious rdfs:labels
• Misdirection: added foaf:depiction triples pointing to replica watches (one triple for each foaf:Person)
• Identity assumption: added malicious owl:sameAs, one for each akt:Organisation
• Inverse-functional property cloning: foaf:Persons cloned and clones are associated same foaf:homepage, considered a IFP.
• Data URI Embedding: added rdfs:seeAlso data:text/html URI for each akt:Organisation
• Malicious subclassing: foaf:Person and akt:Organisation have been defined as subclasses of "GraviaBuyer" and "GraviaSupplier", respectively.