Extracting page shingle hashes for similarity detection.

chrome/renderer/safe_browsing/phishing_term_feature_extractor.cc

Index: chrome/renderer/safe_browsing/phishing_term_feature_extractor.cc
diff --git a/chrome/renderer/safe_browsing/phishing_term_feature_extractor.cc b/chrome/renderer/safe_browsing/phishing_term_feature_extractor.cc
index 89994dfd04cf4488d4f4a87689cff92bc8760bb8..ac53d261d6ca68d7a4059e286cbc7596dde5f71f 100644
--- a/chrome/renderer/safe_browsing/phishing_term_feature_extractor.cc
+++ b/chrome/renderer/safe_browsing/phishing_term_feature_extractor.cc
@@ -45,6 +45,14 @@ struct PhishingTermFeatureExtractor::ExtractionState {
   // Stores up to max_words_per_term_ previous words separated by spaces.
   std::string previous_words;
 
+  // Stores the current shingle after a new word is processed and added in.
+  std::string current_shingle;
+
+  // Stores the sizes of the words in current_shingle. Note: the size includes
+  // the space after each word. In other words, the sum of all sizes in this
+  // list is equal to the length of current_shingle.
+  std::list<size_t> shingle_word_sizes;
+
   // Stores the sizes of the words in previous_words.  Note: the size includes
   // the space after each word.  In other words, the sum of all sizes in this
   // list is equal to the length of previous_words.
@@ -93,13 +101,17 @@ struct PhishingTermFeatureExtractor::ExtractionState {
 PhishingTermFeatureExtractor::PhishingTermFeatureExtractor(
     const base::hash_set<std::string>* page_term_hashes,
     const base::hash_set<uint32>* page_word_hashes,
+    size_t max_hashes_per_page,
     size_t max_words_per_term,
     uint32 murmurhash3_seed,
+    size_t shingle_size,
     FeatureExtractorClock* clock)
     : page_term_hashes_(page_term_hashes),
       page_word_hashes_(page_word_hashes),
+      max_hashes_per_page_(max_hashes_per_page),
       max_words_per_term_(max_words_per_term),
       murmurhash3_seed_(murmurhash3_seed),
+      shingle_size_(shingle_size),
       negative_word_cache_(kMaxNegativeWordCacheSize),
       clock_(clock),
       weak_factory_(this) {
@@ -115,6 +127,7 @@ PhishingTermFeatureExtractor::~PhishingTermFeatureExtractor() {
 void PhishingTermFeatureExtractor::ExtractFeatures(
     const base::string16* page_text,
     FeatureMap* features,
+    std::set<uint32>* shingle_hashes,
     const DoneCallback& done_callback) {
   // The RenderView should have called CancelPendingExtraction() before
   // starting a new extraction, so DCHECK this.
@@ -125,6 +138,7 @@ void PhishingTermFeatureExtractor::ExtractFeatures(
 
   page_text_ = page_text;
   features_ = features;
+  shingle_hashes_ = shingle_hashes,
   done_callback_ = done_callback;
 
   state_.reset(new ExtractionState(*page_text_, clock_->Now()));
@@ -210,6 +224,36 @@ void PhishingTermFeatureExtractor::ExtractFeaturesWithTimeout() {
 
 void PhishingTermFeatureExtractor::HandleWord(
     const base::StringPiece16& word) {
+  // First, extract shingle hashes. We check the size of shingle_hashes_ first
+  // to skip as soon as we reach |max_hashes_per_page_|.
+  std::string word_lower;
+  if (shingle_hashes_->size() < max_hashes_per_page_) {
+    word_lower = base::UTF16ToUTF8(base::i18n::ToLower(word));
+    if (state_->shingle_word_sizes.size() < shingle_size_ - 1) {
+      // Adding |word| would not form a complete shingle, simply insert it.
+      state_->current_shingle.append(word_lower + " ");
+      state_->shingle_word_sizes.push_back(word_lower.size() + 1);
+    } else if (state_->shingle_word_sizes.size() == shingle_size_- 1) {
+      // Adding |word| would just form a complete shingle, insert it and
+      // calculate the shingle hash.
+      state_->current_shingle.append(word_lower + " ");
+      state_->shingle_word_sizes.push_back(word_lower.size() + 1);
+      shingle_hashes_->insert(
+          MurmurHash3String(state_->current_shingle, murmurhash3_seed_));
+    } else {
+      // We need to remove the first word from current_shingle, and add |word|
+      // at the end to form current_shingle.
+      state_->current_shingle.erase(0, state_->shingle_word_sizes.front());
+      state_->shingle_word_sizes.pop_front();
+      state_->current_shingle.append(word_lower + " ");
+      state_->shingle_word_sizes.push_back(word_lower.size() + 1);
+      shingle_hashes_->insert(
+          MurmurHash3String(state_->current_shingle, murmurhash3_seed_));
+    }

[mattm, 2014/05/06 01:00:14]

each of the cases here duplicates the code of the previous case.

You could avoid that by doing:

append word
if (len == shingle_size) {
 calculate & insert shingle
 pop first word
}

[zysxqn, 2014/05/06 20:56:57]

Done.

+  }
+
+  // Next, extract page terms.
+  //
   // Quickest out if we have seen this word before and know that it's not
   // part of any term. This avoids the lowercasing and UTF conversion, both of
   // which are relatively expensive.
@@ -221,7 +265,10 @@ void PhishingTermFeatureExtractor::HandleWord(
     return;
   }
 
-  std::string word_lower = base::UTF16ToUTF8(base::i18n::ToLower(word));
+  // Don't recalculate.
+  if (shingle_hashes_->size() >= max_hashes_per_page_) {

[mattm, 2014/05/06 01:00:14]

I'd probably more comfortable with:
if (word_lower.empty())

[zysxqn, 2014/05/06 20:56:57]

Done.

+    word_lower = base::UTF16ToUTF8(base::i18n::ToLower(word));
+  }
   uint32 word_hash = MurmurHash3String(word_lower, murmurhash3_seed_);
 
   // Quick out if the word is not part of any term, which is the common case.
@@ -302,6 +349,7 @@ void PhishingTermFeatureExtractor::RunCallback(bool success) {
 void PhishingTermFeatureExtractor::Clear() {
   page_text_ = NULL;
   features_ = NULL;
+  shingle_hashes_ = NULL;
   done_callback_.Reset();
   state_.reset(NULL);
   negative_word_cache_.Clear();