Extracting page shingle hashes for similarity detection.

chrome/renderer/safe_browsing/phishing_term_feature_extractor.h

 // Copyright (c) 2011 The Chromium Authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 //
 // PhishingTermFeatureExtractor handles computing term features from the text
 // of a web page for the client-side phishing detection model.  To do this, it
 // takes a list of terms that appear in the model, and scans through the page
 // text looking for them.  Any terms that appear will cause a corresponding
 // features::kPageTerm feature to be added to the FeatureMap.
 //
 // To make it harder for a phisher to enumerate all of the relevant terms in
 // the model, the terms are provided as SHA-256 hashes, rather than plain text.
 //
 // There is one PhishingTermFeatureExtractor per RenderView.
 
 #ifndef CHROME_RENDERER_SAFE_BROWSING_PHISHING_TERM_FEATURE_EXTRACTOR_H_
 #define CHROME_RENDERER_SAFE_BROWSING_PHISHING_TERM_FEATURE_EXTRACTOR_H_
 
+#include <set>
 #include <string>
 
 #include "base/basictypes.h"
 #include "base/callback.h"
 #include "base/containers/hash_tables.h"
 #include "base/containers/mru_cache.h"
 #include "base/memory/scoped_ptr.h"
 #include "base/memory/weak_ptr.h"
 #include "base/strings/string16.h"
 #include "base/strings/string_piece.h"
@@ skipping 11 matching lines @@
   // Creates a PhishingTermFeatureExtractor which will extract features for
   // all of the terms whose SHA-256 hashes are in |page_term_hashes|.  These
   // terms may be multi-word n-grams, with at most |max_words_per_term| words.
   //
   // |page_word_hashes| contains the murmur3 hashes for all of the individual
   // words that make up the terms.  Both sets of strings are UTF-8 encoded and
   // lowercased prior to hashing.  The caller owns both sets of strings, and
   // must ensure that they are valid until the PhishingTermFeatureExtractor is
   // destroyed.
   //
+  // In addition to extracting page terms, we will also extract text shingling
+  // sketch, which consists of hashes of N-gram-words (referred to as shingles)
+  // in the page. |shingle_size| defines N, and |max_shingles_per_page| defines
+  // the maximum number of unique shingle hashes we extracted per page.
+  //
   // |clock| is used for timing feature extractor operations, and may be mocked
   // for testing.  The caller keeps ownership of the clock.
   PhishingTermFeatureExtractor(
       const base::hash_set<std::string>* page_term_hashes,
       const base::hash_set<uint32>* page_word_hashes,
       size_t max_words_per_term,
       uint32 murmurhash3_seed,
+      size_t max_shingles_per_page,
+      size_t shingle_size,
       FeatureExtractorClock* clock);
   ~PhishingTermFeatureExtractor();
 
   // Begins extracting features from |page_text| into the given FeatureMap.
   // |page_text| should contain the plain text of a web page, including any
   // subframes, as returned by RenderView::CaptureText().
   //
   // To avoid blocking the render thread for too long, the feature extractor
   // may run in several chunks of work, posting a task to the current
   // MessageLoop to continue processing.  Once feature extraction is complete,
   // |done_callback| is run on the current thread.
   // PhishingTermFeatureExtractor takes ownership of the callback.
   //
   // |page_text| and |features| are owned by the caller, and must not be
   // destroyed until either |done_callback| is run or
   // CancelPendingExtraction() is called.

[noelutz, 2014/05/06 21:40:18]

nit: mention ownership of shingle_hashes as well?

[zysxqn, 2014/05/07 19:29:19]

Done.

   void ExtractFeatures(const base::string16* page_text,
                        FeatureMap* features,
+                       std::set<uint32>* shingle_hashes,
                        const DoneCallback& done_callback);
 
   // Cancels any pending feature extraction.  The DoneCallback will not be run.
   // Must be called if there is a feature extraction in progress when the page
   // is unloaded or the PhishingTermFeatureExtractor is destroyed.
   void CancelPendingExtraction();
 
  private:
   struct ExtractionState;
 
@@ skipping 43 matching lines @@
   // quick out in the common case that the current word we are processing
   // doesn't contain any part of one of our terms.
   const base::hash_set<uint32>* page_word_hashes_;
 
   // The maximum number of words in an n-gram.
   const size_t max_words_per_term_;
 
   // The seed for murmurhash3.
   const uint32 murmurhash3_seed_;
 
+  // The maximum number of unique shingle hashes we extract in a page.
+  const size_t max_shingles_per_page_;
+
+  // The number of words in a shingle.
+  const size_t shingle_size_;
+
   // This cache is used to see if we need to check the word at all, as
   // converting to UTF8, lowercasing, and hashing are all relatively expensive
   // operations. Though this is called an MRU cache, it seems to behave like
   // an LRU cache (i.e. it evicts the oldest accesses first).
   typedef base::HashingMRUCache<base::StringPiece16, bool> WordCache;
   WordCache negative_word_cache_;
 
   // Non-owned pointer to our clock.
   FeatureExtractorClock* clock_;
 
   // The output parameters from the most recent call to ExtractFeatures().
   const base::string16* page_text_;  // The caller keeps ownership of this.
   FeatureMap* features_;  // The caller keeps ownership of this.
+  std::set<uint32>* shingle_hashes_;
   DoneCallback done_callback_;
 
   // Stores the current state of term extraction from |page_text_|.
   scoped_ptr<ExtractionState> state_;
 
   // Used in scheduling ExtractFeaturesWithTimeout tasks.
   // These pointers are invalidated if extraction is cancelled.
   base::WeakPtrFactory<PhishingTermFeatureExtractor> weak_factory_;
 
   DISALLOW_COPY_AND_ASSIGN(PhishingTermFeatureExtractor);
 };
 
 }  // namespace safe_browsing
 
 #endif  // CHROME_RENDERER_SAFE_BROWSING_PHISHING_TERM_FEATURE_EXTRACTOR_H_