Code/Projects/PacketLevelSignatureExtractor/src/main/java/edu/uci/iotproject/detection/layer3/Layer3ClusterMatcher.java

   1 package edu.uci.iotproject.detection.layer3;
   2
   3 import edu.uci.iotproject.analysis.TriggerTrafficExtractor;
   4 import edu.uci.iotproject.detection.AbstractClusterMatcher;
   5 import edu.uci.iotproject.detection.ClusterMatcherObserver;
   6 import edu.uci.iotproject.trafficreassembly.layer3.Conversation;
   7 import edu.uci.iotproject.trafficreassembly.layer3.TcpReassembler;
   8 import edu.uci.iotproject.analysis.TcpConversationUtils;
   9 import edu.uci.iotproject.io.PcapHandleReader;
  10 import edu.uci.iotproject.util.PrintUtils;
  11 import org.pcap4j.core.*;
  12
  13 import java.time.ZoneId;
  14 import java.util.*;
  15 import java.util.stream.Collectors;
  16
  17 import static edu.uci.iotproject.util.PcapPacketUtils.*;
  18
  19 /**
  20  * Searches a traffic trace for sequences of packets "belong to" a given cluster (in other words, attempts to classify
  21  * traffic as pertaining to a given cluster).
  22  *
  23  * @author Janus Varmarken {@literal <jvarmark@uci.edu>}
  24  * @author Rahmadi Trimananda {@literal <rtrimana@uci.edu>}
  25  */
  26 public class Layer3ClusterMatcher extends AbstractClusterMatcher implements PacketListener {
  27
  28     /**
  29      * The ordered directions of packets in the sequences that make up {@link #mCluster}.
  30      */
  31     private final Conversation.Direction[] mClusterMemberDirections;
  32
  33     /**
  34      * For reassembling the observed traffic into TCP connections.
  35      */
  36     private final TcpReassembler mTcpReassembler = new TcpReassembler();
  37
  38     /**
  39      * IP of the router's WAN port (if analyzed traffic is captured at the ISP's point of view).
  40      */
  41     private final String mRouterWanIp;
  42
  43     /**
  44      * Epsilon value used by the DBSCAN algorithm; it is used again for range-based matching here.
  45      */
  46     private final double mEps;
  47
  48     /**
  49      * The packet inclusion time for signature.
  50      */
  51     private int mInclusionTimeMillis;
  52
  53     /**
  54      * Create a {@link Layer3ClusterMatcher}.
  55      * @param cluster The cluster that traffic is matched against.
  56      * @param routerWanIp The router's WAN IP if examining traffic captured at the ISP's point of view (used for
  57      *                    determining the direction of packets).
  58      * @param inclusionTimeMillis The packet inclusion time for signature.
  59      * @param isRangeBased The boolean that decides if it is range-based vs. strict matching.
  60      * @param eps The epsilon value used in the DBSCAN algorithm.
  61      * @param detectionObservers Client code that wants to get notified whenever the {@link Layer3ClusterMatcher} detects that
  62      *                          (a subset of) the examined traffic is similar to the traffic that makes up
  63      *                          {@code cluster}, i.e., when the examined traffic is classified as pertaining to
  64      *                          {@code cluster}.
  65      */
  66     public Layer3ClusterMatcher(List<List<PcapPacket>> cluster, String routerWanIp, int inclusionTimeMillis,
  67                                 boolean isRangeBased, double eps,
  68                                 ClusterMatcherObserver... detectionObservers) {
  69         super(cluster, isRangeBased);
  70         Objects.requireNonNull(detectionObservers, "detectionObservers cannot be null");
  71         for (ClusterMatcherObserver obs : detectionObservers) {
  72             addObserver(obs);
  73         }
  74         // Build the cluster members' direction sequence.
  75         // Note: assumes that the provided cluster was captured within the local network (routerWanIp is set to null).
  76         mClusterMemberDirections = getPacketDirections(cluster.get(0), null);
  77         /*
  78          * Enforce restriction on cluster members: all representatives must exhibit the same direction pattern and
  79          * contain the same number of packets. Note that this is a somewhat heavy operation, so it may be disabled later
  80          * on in favor of performance. However, it is only run once (at instantiation), so the overhead may be warranted
  81          * in order to ensure correctness, especially during the development/debugging phase.
  82          */
  83         if (!isRangeBased) {    // Only when it is not range-based
  84             if (mCluster.stream().
  85                     anyMatch(inner -> !Arrays.equals(mClusterMemberDirections, getPacketDirections(inner, null)))) {
  86                 throw new IllegalArgumentException(
  87                         "cluster members must contain the same number of packets and exhibit the same packet direction " +
  88                                 "pattern"
  89                 );
  90             }
  91         }
  92         mEps = eps;
  93         mRouterWanIp = routerWanIp;
  94         mInclusionTimeMillis =
  95                 inclusionTimeMillis == 0 ? TriggerTrafficExtractor.INCLUSION_WINDOW_MILLIS : inclusionTimeMillis;
  96     }
  97
  98     @Override
  99     public void gotPacket(PcapPacket packet) {
 100         // Present packet to TCP reassembler so that it can be mapped to a connection (if it is a TCP packet).
 101         mTcpReassembler.gotPacket(packet);
 102     }
 103
 104     /**
 105      * Get the cluster that describes the packet sequence that this {@link Layer3ClusterMatcher} is searching for.
 106      * @return the cluster that describes the packet sequence that this {@link Layer3ClusterMatcher} is searching for.
 107      */
 108     public List<List<PcapPacket>> getCluster() {
 109         return mCluster;
 110     }
 111
 112     public void performDetectionRangeBased() {
 113         /*
 114          * Let's start out simple by building a version that only works for signatures that do not span across multiple
 115          * TCP conversations...
 116          */
 117         for (Conversation c : mTcpReassembler.getTcpConversations()) {
 118             if (c.isTls() && c.getTlsApplicationDataPackets().isEmpty() || !c.isTls() && c.getPackets().isEmpty()) {
 119                 // Skip empty conversations.
 120                 continue;
 121             }
 122             List<PcapPacket> lowerBound = mCluster.get(0);
 123             List<PcapPacket> upperBound = mCluster.get(1);
 124             if (isTlsSequence(lowerBound) != c.isTls() || isTlsSequence(upperBound) != c.isTls()) {
 125                 // We consider it a mismatch if one is a TLS application data sequence and the other is not.
 126                 continue;
 127             }
 128             // Fetch set of packets to examine based on TLS or not.
 129             List<PcapPacket> cPkts = c.isTls() ? c.getTlsApplicationDataPackets() : c.getPackets();
 130             Optional<List<PcapPacket>> match;
 131             while ((match = findSubsequenceInSequence(lowerBound, upperBound, cPkts, mClusterMemberDirections, null)).
 132                     isPresent()) {
 133                 List<PcapPacket> matchSeq = match.get();
 134                 // Notify observers about the match.
 135                 // Max number of skipped packets in layer 3 is 0 (no skipped packets)
 136                 mObservers.forEach(o -> o.onMatch(Layer3ClusterMatcher.this, matchSeq));
 137                 /*
 138                  * Get the index in cPkts of the last packet in the sequence of packets that matches the searched
 139                  * signature sequence.
 140                  */
 141                 int matchSeqEndIdx = cPkts.indexOf(matchSeq.get(matchSeq.size() - 1));
 142                 // We restart the search for the signature sequence immediately after that index, so truncate cPkts.
 143                 cPkts = cPkts.stream().skip(matchSeqEndIdx + 1).collect(Collectors.toList());
 144             }
 145         }
 146     }
 147
 148     public void performDetectionConservative() {
 149         /*
 150          * Let's start out simple by building a version that only works for signatures that do not span across multiple
 151          * TCP conversations...
 152          */
 153         for (Conversation c : mTcpReassembler.getTcpConversations()) {
 154             if (c.isTls() && c.getTlsApplicationDataPackets().isEmpty() || !c.isTls() && c.getPackets().isEmpty()) {
 155                 // Skip empty conversations.
 156                 continue;
 157             }
 158             for (List<PcapPacket> signatureSequence : mCluster) {
 159                 if (isTlsSequence(signatureSequence) != c.isTls()) {
 160                     // We consider it a mismatch if one is a TLS application data sequence and the other is not.
 161                     continue;
 162                 }
 163                 // Fetch set of packets to examine based on TLS or not.
 164                 List<PcapPacket> cPkts = c.isTls() ? c.getTlsApplicationDataPackets() : c.getPackets();
 165                 /*
 166                  * Note: we embed the attempt to detect the signature sequence in a loop in order to capture those cases
 167                  * where the same signature sequence appears multiple times in one Conversation.
 168                  *
 169                  * Note: since we expect all sequences that together make up the signature to exhibit the same direction
 170                  * pattern, we can simply pass the precomputed direction array for the signature sequence so that it
 171                  * won't have to be recomputed internally in each call to findSubsequenceInSequence().
 172                  */
 173                 Optional<List<PcapPacket>> match;
 174                 while ((match = findSubsequenceInSequence(signatureSequence, cPkts, mClusterMemberDirections, null)).
 175                         isPresent()) {
 176                     List<PcapPacket> matchSeq = match.get();
 177                     // Notify observers about the match.
 178                     // Max number of skipped packets in layer 3 is 0 (no skipped packets)
 179                     mObservers.forEach(o -> o.onMatch(Layer3ClusterMatcher.this, matchSeq));
 180                     /*
 181                      * Get the index in cPkts of the last packet in the sequence of packets that matches the searched
 182                      * signature sequence.
 183                      */
 184                     int matchSeqEndIdx = cPkts.indexOf(matchSeq.get(matchSeq.size() - 1));
 185                     // We restart the search for the signature sequence immediately after that index, so truncate cPkts.
 186                     cPkts = cPkts.stream().skip(matchSeqEndIdx + 1).collect(Collectors.toList());
 187                 }
 188             }
 189
 190             /*
 191              * TODO:
 192              * if no item in cluster matches, also perform a distance-based matching to cover those cases where we did
 193              * not manage to capture every single mutation of the sequence during training.
 194              *
 195              * Need to compute average/centroid of cluster to do so...? Compute within-cluster variance, then check if
 196              * distance between input conversation and cluster average/centroid is smaller than or equal to the computed
 197              * variance?
 198              */
 199         }
 200     }
 201
 202     /**
 203      * Checks if {@code sequence} is a sequence of TLS packets. Note: the current implementation relies on inspection
 204      * of the port numbers when deciding between TLS vs. non-TLS. Therefore, only the first packet of {@code sequence}
 205      * is examined as it is assumed that all packets in {@code sequence} pertain to the same {@link Conversation} and
 206      * hence share the same set of two src/dst port numbers (albeit possibly alternating between which one is the src
 207      * and which one is the dst, as packets in {@code sequence} may be in alternating directions).
 208      * @param sequence The sequence of packets for which it is to be determined if it is a sequence of TLS packets or
 209      *                 non-TLS packets.
 210      * @return {@code true} if {@code sequence} is a sequence of TLS packets, {@code false} otherwise.
 211      */
 212     private boolean isTlsSequence(List<PcapPacket> sequence) {
 213         // NOTE: Assumes ALL packets in sequence pertain to the same TCP connection!
 214         PcapPacket firstPkt = sequence.get(0);
 215         int srcPort = getSourcePort(firstPkt);
 216         int dstPort = getDestinationPort(firstPkt);
 217         return TcpConversationUtils.isTlsPort(srcPort) || TcpConversationUtils.isTlsPort(dstPort);
 218     }
 219
 220     /**
 221      * Examine if a given sequence of packets ({@code sequence}) contains a given shorter sequence of packets
 222      * ({@code subsequence}). Note: the current implementation actually searches for a substring as it does not allow
 223      * for interleaving packets in {@code sequence} that are not in {@code subsequence}; for example, if
 224      * {@code subsequence} consists of packet lengths [2, 3, 5] and {@code sequence} consists of  packet lengths
 225      * [2, 3, 4, 5], the result will be that there is no match (because of the interleaving 4). If we are to allow
 226      * interleaving packets, we need a modified version of
 227      * <a href="https://stackoverflow.com/a/20545604/1214974">this</a>.
 228      *
 229      * @param subsequence The sequence to search for.
 230      * @param sequence The sequence to search.
 231      * @param subsequenceDirections The directions of packets in {@code subsequence} such that for all {@code i},
 232      *                              {@code subsequenceDirections[i]} is the direction of the packet returned by
 233      *                              {@code subsequence.get(i)}. May be set to {@code null}, in which this call will
 234      *                              internally compute the packet directions.
 235      * @param sequenceDirections The directions of packets in {@code sequence} such that for all {@code i},
 236      *                           {@code sequenceDirections[i]} is the direction of the packet returned by
 237      *                           {@code sequence.get(i)}. May be set to {@code null}, in which this call will internally
 238      *                           compute the packet directions.
 239      *
 240      * @return An {@link Optional} containing the part of {@code sequence} that matches {@code subsequence}, or an empty
 241      *         {@link Optional} if no part of {@code sequence} matches {@code subsequence}.
 242      */
 243     private Optional<List<PcapPacket>> findSubsequenceInSequence(List<PcapPacket> subsequence,
 244                                                                  List<PcapPacket> sequence,
 245                                                                  Conversation.Direction[] subsequenceDirections,
 246                                                                  Conversation.Direction[] sequenceDirections) {
 247         if (sequence.size() < subsequence.size()) {
 248             // If subsequence is longer, it cannot be contained in sequence.
 249             return Optional.empty();
 250         }
 251         if (isTlsSequence(subsequence) != isTlsSequence(sequence)) {
 252             // We consider it a mismatch if one is a TLS application data sequence and the other is not.
 253             return Optional.empty();
 254         }
 255         // If packet directions have not been precomputed by calling code, we need to construct them.
 256         if (subsequenceDirections == null) {
 257             subsequenceDirections = getPacketDirections(subsequence, mRouterWanIp);
 258         }
 259         if (sequenceDirections == null) {
 260             sequenceDirections = getPacketDirections(sequence, mRouterWanIp);
 261         }
 262         int subseqIdx = 0;
 263         int seqIdx = 0;
 264         while (seqIdx < sequence.size()) {
 265             PcapPacket subseqPkt = subsequence.get(subseqIdx);
 266             PcapPacket seqPkt = sequence.get(seqIdx);
 267             // We only have a match if packet lengths and directions match.
 268             if (subseqPkt.getOriginalLength() == seqPkt.getOriginalLength() &&
 269                     subsequenceDirections[subseqIdx] == sequenceDirections[seqIdx]) {
 270                 // A match; advance both indices to consider next packet in subsequence vs. next packet in sequence.
 271                 subseqIdx++;
 272                 seqIdx++;
 273                 if (subseqIdx == subsequence.size()) {
 274                     // We managed to match the entire subsequence in sequence.
 275                     // Return the sublist of sequence that matches subsequence.
 276                     /*
 277                      * TODO:
 278                      * ASSUMES THE BACKING LIST (i.e., 'sequence') IS _NOT_ STRUCTURALLY MODIFIED, hence may not work
 279                      * for live traces!
 280                      */
 281                     return Optional.of(sequence.subList(seqIdx - subsequence.size(), seqIdx));
 282                 }
 283             } else {
 284                 // Mismatch.
 285                 if (subseqIdx > 0) {
 286                     /*
 287                      * If we managed to match parts of subsequence, we restart the search for subsequence in sequence at
 288                      * the index of sequence where the current mismatch occurred. I.e., we must reset subseqIdx, but
 289                      * leave seqIdx untouched.
 290                      */
 291                     subseqIdx = 0;
 292                 } else {
 293                     /*
 294                      * First packet of subsequence didn't match packet at seqIdx of sequence, so we move forward in
 295                      * sequence, i.e., we continue the search for subsequence in sequence starting at index seqIdx+1 of
 296                      * sequence.
 297                      */
 298                     seqIdx++;
 299                 }
 300             }
 301         }
 302         return Optional.empty();
 303     }
 304
 305     /**
 306      * Overloading the method {@code findSubsequenceInSequence} for range-based matching. Instead of a sequence,
 307      * we have sequences of lower and upper bounds.
 308      *
 309      * @param lowerBound The lower bound of the sequence we search for.
 310      * @param upperBound The upper bound of the sequence we search for.
 311      * @param subsequenceDirections The directions of packets in {@code subsequence} such that for all {@code i},
 312      *                              {@code subsequenceDirections[i]} is the direction of the packet returned by
 313      *                              {@code subsequence.get(i)}. May be set to {@code null}, in which this call will
 314      *                              internally compute the packet directions.
 315      * @param sequenceDirections The directions of packets in {@code sequence} such that for all {@code i},
 316      *                           {@code sequenceDirections[i]} is the direction of the packet returned by
 317      *                           {@code sequence.get(i)}. May be set to {@code null}, in which this call will internally
 318      *                           compute the packet directions.
 319      *
 320      * @return An {@link Optional} containing the part of {@code sequence} that matches {@code subsequence}, or an empty
 321      *         {@link Optional} if no part of {@code sequence} matches {@code subsequence}.
 322      */
 323     private Optional<List<PcapPacket>> findSubsequenceInSequence(List<PcapPacket> lowerBound,
 324                                                                  List<PcapPacket> upperBound,
 325                                                                  List<PcapPacket> sequence,
 326                                                                  Conversation.Direction[] subsequenceDirections,
 327                                                                  Conversation.Direction[] sequenceDirections) {
 328         // Just do the checks for either lower or upper bound!
 329         // TODO: For now we use just the lower bound
 330         if (sequence.size() < lowerBound.size()) {
 331             // If subsequence is longer, it cannot be contained in sequence.
 332             return Optional.empty();
 333         }
 334         if (isTlsSequence(lowerBound) != isTlsSequence(sequence)) {
 335             // We consider it a mismatch if one is a TLS application data sequence and the other is not.
 336             return Optional.empty();
 337         }
 338         // If packet directions have not been precomputed by calling code, we need to construct them.
 339         if (subsequenceDirections == null) {
 340             subsequenceDirections = getPacketDirections(lowerBound, mRouterWanIp);
 341         }
 342         if (sequenceDirections == null) {
 343             sequenceDirections = getPacketDirections(sequence, mRouterWanIp);
 344         }
 345         int subseqIdx = 0;
 346         int seqIdx = 0;
 347         while (seqIdx < sequence.size()) {
 348             PcapPacket lowBndPkt = lowerBound.get(subseqIdx);
 349             PcapPacket upBndPkt = upperBound.get(subseqIdx);
 350             PcapPacket seqPkt = sequence.get(seqIdx);
 351             // We only have a match if packet lengths and directions match.
 352             // The packet lengths have to be in the range of [lowerBound - eps, upperBound+eps]
 353             // We initialize the lower and upper bounds first
 354             int epsLowerBound = lowBndPkt.length();
 355             int epsUpperBound = upBndPkt.length();
 356             // Do strict matching if the lower and upper bounds are the same length
 357             // Do range matching with eps otherwise
 358             if (epsLowerBound != epsUpperBound) {
 359                 // TODO: Maybe we could do better here for the double to integer conversion?
 360                 epsLowerBound = epsLowerBound - (int) mEps;
 361                 epsUpperBound = epsUpperBound + (int) mEps;
 362             }
 363             if (epsLowerBound <= seqPkt.getOriginalLength() &&
 364                     seqPkt.getOriginalLength() <= epsUpperBound &&
 365                     subsequenceDirections[subseqIdx] == sequenceDirections[seqIdx]) {
 366                 // A match; advance both indices to consider next packet in subsequence vs. next packet in sequence.
 367                 subseqIdx++;
 368                 seqIdx++;
 369                 if (subseqIdx == lowerBound.size()) {
 370                     // We managed to match the entire subsequence in sequence.
 371                     // Return the sublist of sequence that matches subsequence.
 372                     /*
 373                      * TODO:
 374                      * ASSUMES THE BACKING LIST (i.e., 'sequence') IS _NOT_ STRUCTURALLY MODIFIED, hence may not work
 375                      * for live traces!
 376                      */
 377                     return Optional.of(sequence.subList(seqIdx - lowerBound.size(), seqIdx));
 378                 }
 379             } else {
 380                 // Mismatch.
 381                 if (subseqIdx > 0) {
 382                     /*
 383                      * If we managed to match parts of subsequence, we restart the search for subsequence in sequence at
 384                      * the index of sequence where the current mismatch occurred. I.e., we must reset subseqIdx, but
 385                      * leave seqIdx untouched.
 386                      */
 387                     subseqIdx = 0;
 388                 } else {
 389                     /*
 390                      * First packet of subsequence didn't match packet at seqIdx of sequence, so we move forward in
 391                      * sequence, i.e., we continue the search for subsequence in sequence starting at index seqIdx+1 of
 392                      * sequence.
 393                      */
 394                     seqIdx++;
 395                 }
 396             }
 397         }
 398         return Optional.empty();
 399     }
 400
 401     /**
 402      * Given a cluster, produces a pruned version of that cluster. In the pruned version, there are no duplicate cluster
 403      * members. Two cluster members are considered identical if their packets lengths and packet directions are
 404      * identical. The resulting pruned cluster is unmodifiable (this applies to both the outermost list as well as the
 405      * nested lists) in order to preserve its integrity when exposed to external code (e.g., through
 406      * {@link #getCluster()}).
 407      *
 408      * @param cluster A cluster to prune.
 409      * @return The resulting pruned cluster.
 410      */
 411     @Override
 412     protected List<List<PcapPacket>> pruneCluster(List<List<PcapPacket>> cluster) {
 413         List<List<PcapPacket>> prunedCluster = new ArrayList<>();
 414         for (List<PcapPacket> originalClusterSeq : cluster) {
 415             boolean alreadyPresent = false;
 416             for (List<PcapPacket> prunedClusterSeq : prunedCluster) {
 417                 Optional<List<PcapPacket>> duplicate = findSubsequenceInSequence(originalClusterSeq, prunedClusterSeq,
 418                         mClusterMemberDirections, mClusterMemberDirections);
 419                 if (duplicate.isPresent()) {
 420                     alreadyPresent = true;
 421                     break;
 422                 }
 423             }
 424             if (!alreadyPresent) {
 425                 prunedCluster.add(Collections.unmodifiableList(originalClusterSeq));
 426             }
 427         }
 428         return Collections.unmodifiableList(prunedCluster);
 429     }
 430
 431     /**
 432      * Given a {@code List<PcapPacket>}, generate a {@code Conversation.Direction[]} such that each entry in the
 433      * resulting {@code Conversation.Direction[]} specifies the direction of the {@link PcapPacket} at the corresponding
 434      * index in the input list.
 435      * @param packets The list of packets for which to construct a corresponding array of packet directions.
 436      * @param routerWanIp The IP of the router's WAN port. This is used for determining the direction of packets when
 437      *                    the traffic is captured just outside the local network (at the ISP side of the router). Set to
 438      *                    {@code null} if {@code packets} stem from traffic captured within the local network.
 439      * @return A {@code Conversation.Direction[]} specifying the direction of the {@link PcapPacket} at the
 440      *         corresponding index in {@code packets}.
 441      */
 442     private static Conversation.Direction[] getPacketDirections(List<PcapPacket> packets, String routerWanIp) {
 443         Conversation.Direction[] directions = new Conversation.Direction[packets.size()];
 444         for (int i = 0; i < packets.size(); i++) {
 445             PcapPacket pkt = packets.get(i);
 446             if (getSourceIp(pkt).equals(getDestinationIp(pkt))) {
 447                 // Sanity check: we shouldn't be processing loopback traffic
 448                 throw new AssertionError("loopback traffic detected");
 449             }
 450             if (isSrcIpLocal(pkt) || getSourceIp(pkt).equals(routerWanIp)) {
 451                 directions[i] = Conversation.Direction.CLIENT_TO_SERVER;
 452             } else if (isDstIpLocal(pkt) || getDestinationIp(pkt).equals(routerWanIp)) {
 453                 directions[i] = Conversation.Direction.SERVER_TO_CLIENT;
 454             } else {
 455                 //throw new IllegalArgumentException("no local IP or router WAN port IP found, can't detect direction");
 456             }
 457         }
 458         return directions;
 459     }
 460
 461 }