7 Preventing background knowledge attacks 8 Conclusions Permutation-based Sequential Pattern Hiding

Sequence data increasingly shared by organizations to enable mining applications market-basket analysis web and process modeling preference-based services Mining may expose sensitive sequential patterns, which allow intrusive inferences to individuals provide competitive advance to organizations This threat causes unsolicited advertisement, customer profiling, financial and reputational loss, etc. It has transpired in several cases (e.g., Walmart- P&G). Permutation-based Sequential Pattern Hiding 4Motivation (2/3) Insurance company shares sequence data with marketing partners. Data can be mined to find the number of customers who bought sports car, travel, and home insurance in this order ([f, g, b]), or travel, home, and trailer insurance in any order ({g, b, d}). Sensitive patterns are exposed when their support is at least 3. e.g., [a, c, e] → insurance purchase for motorbike, jet-ski and then truck a b c d e f a b c e a c e h b f g c e a b d f g h b d h b f g f h g b c d f g e d h f g b Original data [a, c, e] [d, f, g] [d, h, b] Sensitive patterns To preserve privacy and mining accuracy Upper-bound the support of sensitive patterns (minSup threshold), so that they are no longer actionable. Minimize side-effects (changes in frequent, nonsensitive patterns) Permutation-based Sequential Pattern Hiding 5Motivation (3/3) All existing methods [1, 3] use symbol deletion. minimal number of deleted symbols and/or side-effects Deletion is a drastic operation- it reduces symbol support utility loss in sequence mining and tasks based on itemset properties. The state-of-the-art algorithm, SBSH [3], applied with minSup = 3: 3 symbol deletions, 5 side-effects (e.g., [a, c] is no longer frequent) 12 out of 48 frequent itemsets are not mineable (lost) a b c d e f * b c e a c e h b f g c e a b d f g h b * h b f g f h g b c * f g e d h f g