chrome/renderer/safe_browsing/phishing_classifier.cc

   1 // Copyright (c) 2011 The Chromium Authors. All rights reserved.
   2 // Use of this source code is governed by a BSD-style license that can be
   3 // found in the LICENSE file.
   4
   5 #include "chrome/renderer/safe_browsing/phishing_classifier.h"
   6
   7 #include <string>
   8
   9 #include "base/bind.h"
  10 #include "base/callback.h"
  11 #include "base/compiler_specific.h"
  12 #include "base/logging.h"
  13 #include "base/message_loop/message_loop.h"
  14 #include "base/metrics/histogram.h"
  15 #include "base/strings/string_util.h"
  16 #include "chrome/common/safe_browsing/csd.pb.h"
  17 #include "chrome/common/url_constants.h"
  18 #include "chrome/renderer/safe_browsing/feature_extractor_clock.h"
  19 #include "chrome/renderer/safe_browsing/features.h"
  20 #include "chrome/renderer/safe_browsing/phishing_dom_feature_extractor.h"
  21 #include "chrome/renderer/safe_browsing/phishing_term_feature_extractor.h"
  22 #include "chrome/renderer/safe_browsing/phishing_url_feature_extractor.h"
  23 #include "chrome/renderer/safe_browsing/scorer.h"
  24 #include "content/public/renderer/render_view.h"
  25 #include "crypto/sha2.h"
  26 #include "third_party/WebKit/public/platform/WebURL.h"
  27 #include "third_party/WebKit/public/platform/WebURLRequest.h"
  28 #include "third_party/WebKit/public/web/WebDataSource.h"
  29 #include "third_party/WebKit/public/web/WebDocument.h"
  30 #include "third_party/WebKit/public/web/WebFrame.h"
  31 #include "third_party/WebKit/public/web/WebView.h"
  32 #include "url/gurl.h"
  33
  34 namespace safe_browsing {
  35
  36 const float PhishingClassifier::kInvalidScore = -1.0;
  37 const float PhishingClassifier::kPhishyThreshold = 0.5;
  38
  39 PhishingClassifier::PhishingClassifier(content::RenderView* render_view,
  40                                        FeatureExtractorClock* clock)
  41     : render_view_(render_view),
  42       scorer_(NULL),
  43       clock_(clock),
  44       weak_factory_(this) {
  45   Clear();
  46 }
  47
  48 PhishingClassifier::~PhishingClassifier() {
  49   // The RenderView should have called CancelPendingClassification() before
  50   // we are destroyed.
  51   CheckNoPendingClassification();
  52 }
  53
  54 void PhishingClassifier::set_phishing_scorer(const Scorer* scorer) {
  55   CheckNoPendingClassification();
  56   scorer_ = scorer;
  57   if (scorer_) {
  58     url_extractor_.reset(new PhishingUrlFeatureExtractor);
  59     dom_extractor_.reset(
  60         new PhishingDOMFeatureExtractor(render_view_, clock_.get()));
  61     term_extractor_.reset(new PhishingTermFeatureExtractor(
  62         &scorer_->page_terms(),
  63         &scorer_->page_words(),
  64         scorer_->max_words_per_term(),
  65         scorer_->murmurhash3_seed(),
  66         scorer_->max_shingles_per_page(),
  67         scorer_->shingle_size(),
  68         clock_.get()));
  69   } else {
  70     // We're disabling client-side phishing detection, so tear down all
  71     // of the relevant objects.
  72     url_extractor_.reset();
  73     dom_extractor_.reset();
  74     term_extractor_.reset();
  75   }
  76 }
  77
  78 bool PhishingClassifier::is_ready() const {
  79   return scorer_ != NULL;
  80 }
  81
  82 void PhishingClassifier::BeginClassification(
  83     const base::string16* page_text,
  84     const DoneCallback& done_callback) {
  85   DCHECK(is_ready());
  86
  87   // The RenderView should have called CancelPendingClassification() before
  88   // starting a new classification, so DCHECK this.
  89   CheckNoPendingClassification();
  90   // However, in an opt build, we will go ahead and clean up the pending
  91   // classification so that we can start in a known state.
  92   CancelPendingClassification();
  93
  94   page_text_ = page_text;
  95   done_callback_ = done_callback;
  96
  97   // For consistency, we always want to invoke the DoneCallback
  98   // asynchronously, rather than directly from this method.  To ensure that
  99   // this is the case, post a task to begin feature extraction on the next
 100   // iteration of the message loop.
 101   base::MessageLoop::current()->PostTask(
 102       FROM_HERE,
 103       base::Bind(&PhishingClassifier::BeginFeatureExtraction,
 104                  weak_factory_.GetWeakPtr()));
 105 }
 106
 107 void PhishingClassifier::BeginFeatureExtraction() {
 108   blink::WebView* web_view = render_view_->GetWebView();
 109   if (!web_view) {
 110     RunFailureCallback();
 111     return;
 112   }
 113
 114   blink::WebFrame* frame = web_view->mainFrame();
 115   if (!frame) {
 116     RunFailureCallback();
 117     return;
 118   }
 119
 120   // Check whether the URL is one that we should classify.
 121   // Currently, we only classify http: URLs that are GET requests.
 122   GURL url(frame->document().url());
 123   if (!url.SchemeIs(url::kHttpScheme)) {
 124     RunFailureCallback();
 125     return;
 126   }
 127
 128   blink::WebDataSource* ds = frame->dataSource();
 129   if (!ds || !EqualsASCII(ds->request().httpMethod(), "GET")) {
 130     RunFailureCallback();
 131     return;
 132   }
 133
 134   features_.reset(new FeatureMap);
 135   if (!url_extractor_->ExtractFeatures(url, features_.get())) {
 136     RunFailureCallback();
 137     return;
 138   }
 139
 140   // DOM feature extraction can take awhile, so it runs asynchronously
 141   // in several chunks of work and invokes the callback when finished.
 142   dom_extractor_->ExtractFeatures(
 143       features_.get(),
 144       base::Bind(&PhishingClassifier::DOMExtractionFinished,
 145                  base::Unretained(this)));
 146 }
 147
 148 void PhishingClassifier::CancelPendingClassification() {
 149   // Note that cancelling the feature extractors is simply a no-op if they
 150   // were not running.
 151   DCHECK(is_ready());
 152   dom_extractor_->CancelPendingExtraction();
 153   term_extractor_->CancelPendingExtraction();
 154   weak_factory_.InvalidateWeakPtrs();
 155   Clear();
 156 }
 157
 158 void PhishingClassifier::DOMExtractionFinished(bool success) {
 159   shingle_hashes_.reset(new std::set<uint32>);
 160   if (success) {
 161     // Term feature extraction can take awhile, so it runs asynchronously
 162     // in several chunks of work and invokes the callback when finished.
 163     term_extractor_->ExtractFeatures(
 164         page_text_,
 165         features_.get(),
 166         shingle_hashes_.get(),
 167         base::Bind(&PhishingClassifier::TermExtractionFinished,
 168                    base::Unretained(this)));
 169   } else {
 170     RunFailureCallback();
 171   }
 172 }
 173
 174 void PhishingClassifier::TermExtractionFinished(bool success) {
 175   if (success) {
 176     blink::WebView* web_view = render_view_->GetWebView();
 177     if (!web_view) {
 178       RunFailureCallback();
 179       return;
 180     }
 181     blink::WebFrame* main_frame = web_view->mainFrame();
 182     if (!main_frame) {
 183       RunFailureCallback();
 184       return;
 185     }
 186
 187     // Hash all of the features so that they match the model, then compute
 188     // the score.
 189     FeatureMap hashed_features;
 190     ClientPhishingRequest verdict;
 191     verdict.set_model_version(scorer_->model_version());
 192     verdict.set_url(main_frame->document().url().spec());
 193     for (base::hash_map<std::string, double>::const_iterator it =
 194              features_->features().begin();
 195          it != features_->features().end(); ++it) {
 196       DVLOG(2) << "Feature: " << it->first << " = " << it->second;
 197       bool result = hashed_features.AddRealFeature(
 198           crypto::SHA256HashString(it->first), it->second);
 199       DCHECK(result);
 200       ClientPhishingRequest::Feature* feature = verdict.add_feature_map();
 201       feature->set_name(it->first);
 202       feature->set_value(it->second);
 203     }
 204     for (std::set<uint32>::const_iterator it = shingle_hashes_->begin();
 205          it != shingle_hashes_->end(); ++it) {
 206       verdict.add_shingle_hashes(*it);
 207     }
 208     float score = static_cast<float>(scorer_->ComputeScore(hashed_features));
 209     verdict.set_client_score(score);
 210     verdict.set_is_phishing(score >= kPhishyThreshold);
 211     RunCallback(verdict);
 212   } else {
 213     RunFailureCallback();
 214   }
 215 }
 216
 217 void PhishingClassifier::CheckNoPendingClassification() {
 218   DCHECK(done_callback_.is_null());
 219   DCHECK(!page_text_);
 220   if (!done_callback_.is_null() || page_text_) {
 221     LOG(ERROR) << "Classification in progress, missing call to "
 222                << "CancelPendingClassification";
 223     UMA_HISTOGRAM_COUNTS("SBClientPhishing.CheckNoPendingClassificationFailed",
 224                          1);
 225   }
 226 }
 227
 228 void PhishingClassifier::RunCallback(const ClientPhishingRequest& verdict) {
 229   done_callback_.Run(verdict);
 230   Clear();
 231 }
 232
 233 void PhishingClassifier::RunFailureCallback() {
 234   ClientPhishingRequest verdict;
 235   // In this case we're not guaranteed to have a valid URL.  Just set it
 236   // to the empty string to make sure we have a valid protocol buffer.
 237   verdict.set_url("");
 238   verdict.set_client_score(kInvalidScore);
 239   verdict.set_is_phishing(false);
 240   RunCallback(verdict);
 241 }
 242
 243 void PhishingClassifier::Clear() {
 244   page_text_ = NULL;
 245   done_callback_.Reset();
 246   features_.reset(NULL);
 247   shingle_hashes_.reset(NULL);
 248 }
 249
 250 }  // namespace safe_browsing