Advanced Classification Lists (Dirty Word Lists) for Automatic Security Classification

Automatic security classification by machine learning for cross-domain information exchange

“…Among different contributions, this paper investigates the effects of topic classification prior to security classification, as only assumed in previous works [11], [12]. To the best of our knowledge, the works in [8], [9] and [10] were the first published papers on automatic security classification and the application of machine learning techniques to this problem.…”

“…The Digital National Security Archive contains the most comprehensive collection of declassified US government documents available to the public [21]. We base our analysis on documents from the same collection that were used in [8], [9] and [10]. This collection was originally chosen because it contains a mix of both classified and unclassified documents from three unrelated domains: Years, 1965-1986 Of the 5853 documents available within these three topics, we do not use duplicate documents and documents classes that are very small or have an unsuitable classification [8], [9].…”

“…We base our analysis on documents from the same collection that were used in [8], [9] and [10]. This collection was originally chosen because it contains a mix of both classified and unclassified documents from three unrelated domains: Years, 1965-1986 Of the 5853 documents available within these three topics, we do not use duplicate documents and documents classes that are very small or have an unsuitable classification [8], [9]. Then we remove documents with 30 words or less (after having also removed some keywords as explained below), and we end up with 2805 documents in total (Table I).…”

“…The first class is Unclassified, which comprises 1079 documents. The second class is Classified, which is an aggregation of the document classes "Confidential", "Secret" and "Top Secret" in Table I [8], [9] [10]. However, then we remove some of the latter documents to get an approximately equal share between classes, ending up with around 2158 documents.…”

“…The contribution of this paper is putting this topic on the agenda, and exploring techniques for content analysis that is more sophisticated than the dirty word scanning techniques that are currently deployed. In this paper, we employ a machine learning approach to the problem, and to the best of our knowledge, this paper is together with our previous works in [8], [9] and [10] the only published work addressing this issue. Other relevant available information we have found is a Master thesis from 2008 [11], as well as a military (thus, difficultly accessible) technical report [12].…”