Electrical computers and digital processing systems: support – Multiple computer communication using cryptography – Protection at a particular protocol layer
Reexamination Certificate
1999-10-25
2004-07-27
Wright, Norman M. (Department: 2134)
Electrical computers and digital processing systems: support
Multiple computer communication using cryptography
Protection at a particular protocol layer
C713S152000, C700S079000, C700S110000
Reexamination Certificate
active
06769066
ABSTRACT:
BACKGROUND OF THE INVENTION
1. Field of the Invention
The present invention relates generally to the field of computer systems software and computer network security. More specifically, it relates to software for examining user and group activity in a computer network and for training a model for use in detecting potential security violations in the network.
2. Discussion of Related Art
Computer network security is an important issue for all types of organizations and enterprises. Computer break-ins and their misuse have become common features. The number, as well as sophistication, of attacks on computer systems is on the rise. Often, network intruders have easily overcome the password authentication mechanism designed to protect the system. With an increased understanding of how systems work, intruders have become skilled at determining their weaknesses and exploiting them to obtain unauthorized privileges. Intruders also use patterns of intrusion that are often difficult to trace and identify. They use several levels of indirection before breaking into target systems and rarely indulge in sudden bursts of suspicious or anomalous activity. If an account on a target system is compromised, intruders can carefully cover their tracks as not to arouse suspicion. Furthermore, threats like viruses and worms do not need human supervision and are capable of replicating and traveling to connected computer systems. Unleashed at one computer, by the time they are discovered, it is almost impossible to trace their origin or the extent of infection.
As the number of users within a particular entity grows, the risks from unauthorized intrusions into computer systems or into certain sensitive components of a large computer system increase. In order to maintain a reliable and secure computer network, regardless of network size, exposure to potential network intrusions must be reduced as much as possible. Network intrusions can originate from legitimate users within an entity attempting to access secure portions of the network or can originate from illegitimate users outside an entity attempting to break into the entity's network often referred to as “hackers.” Intrusions from either of these two groups of users can be damaging to an organization's computer network. Most attempted security violations are internal; that is, they are attempted by employees of an enterprise or organization.
One approach to detecting computer network intrusions is calculating “features” based on various factors, such as command sequences, user activity, machine usage loads, and resource violations, files accessed, data transferred, terminal activity, network activity, among others. Features are then used as input to a model or expert system which determines whether a possible intrusion or violation has occurred. The use of features is well-known in various fields in computer science including the field of computer network security, especially in conjunction with an expert system which evaluates the feature values. Features used in present computer security systems are generally rule-based features. Such features lead to computer security systems that are inflexible, highly complex, and require frequent upgrading and maintenance.
Expert systems that use such features generally use thresholds (e.g., “if-then-else” clauses, “case” statements, etc.) to determine whether there was a violation. Thus, a human expert with extensive knowledge of the computer network domain has to accurately determine and assign such thresholds for the system to be effective. These thresholds and other rules are typically not modified often and do not reflect day-to-day fluctuations based on changing user behavior. Such rules are typically entered by an individual with extensive domain knowledge of the particular system. In short, such systems lack the robustness needed to detect increasingly sophisticated lines of attack in a computer system. A reliable computer system must be able to accurately determine when a possible intrusion is occurring and who the intruder is, and do so by taking into account trends in user activity.
As mentioned above, rule-based features can also be used as input to a model instead of an expert system. However, a model that can accept only rule-based features and cannot be trained to adjust to trends and changing needs in a computer network generally suffers from the same drawbacks as the expert system configuration. A model is generally used in conjunction with a features generator and accepts as input a features list. However, models presently used in computer network intrusion detection systems are not trained to take into account changing requirements and user trends in a computer network. Thus, such models also lead to computer security systems that are inflexible, complex, and require frequent upgrading and maintenance.
FIG. 1
is a block diagram depicting certain components in a security system in a computer network as is presently known in the art. A features/expert systems component
10
of a complete network security system (not shown) has three general components: user activity
12
, expert system
14
, and alert messages
16
. User activity
12
contains “raw” data, typically in the form of aggregated log files and is raw in that it is typically unmodified or has not gone through significant preprocessing. User activity
12
has records of actions taken by users on the network that the organization or enterprise wants to monitor.
Expert system
14
, also referred to as a “rule-based” engine, accepts input data from user activity files
12
which acts as features in present security systems. As mentioned above, the expert system, a term well-understood in the field of computer science, processes the input features and determines, based on its rules, whether a violation has occurred or whether there is anomalous activity. In two simple examples, expert system
14
can contain a rule instructing it to issue an alert message if a user attempts to logon using an incorrect password more than five consecutive times or if a user attempts to write to a restricted file more than once.
Alert message
16
is issued if a rule threshold is exceeded to inform a network security analyst that a possible intrusion may be occurring. Typically, alert message
16
contains a score and a reason for the alert, i.e., which rules or thresholds were violated by a user. As stated above, these thresholds can be outdated or moot if circumstances change in the system. For example, circumstances can change and the restricted file mentioned above can be made accessible to a larger group of users. In this case an expert would have to modify the rules in expert system
14
.
As mentioned above, the feature and expert system components as shown in FIG.
1
and conventional models used in conjunction with these components have significant drawbacks. One is the cumbersome and overly complex set of rules and thresholds that must be entered to “cover” all the possible security violations. Another is the knowledge an expert must have in order to update or modify the rule base and the model to reflect changing circumstances in the organization. Related to this is the difficulty in locating an expert to assist in programming and maintaining all components in the system.
Therefore, it would be desirable to utilize a features list generator in place of a traditional expert system that can automatically update itself to reflect changes in user and user group current behavior. It would also be desirable to derive a training process for a model used in conjunction with a features generator to generate a score reflective of changing user behavior. It would also be desirable to have the training process or algorithm accurately read anomalous user behavior. Furthermore, it would be desirable to have such a features generator be self-sufficient and flexible in that it is not dependent on changes entered by an expert and is not a rigid rule-based system.
SUMMARY OF THE INVENTION
To achieve the foregoing, methods, apparatus, and computer-read
Botros Sherif M.
Diep Thanh A.
Izenson Martin D.
Beyer Weaver & Thomas LLP
Visa International Service Association
Wright Norman M.
LandOfFree
Method and apparatus for training a neural network model for... does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Method and apparatus for training a neural network model for..., we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Method and apparatus for training a neural network model for... will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3236518