Artificial Immune Systems AIS Model Tutorial

2005-12-13 Y. Tan---Artificial Immune Sys. 1

Artificial Immune System and Its Applications

Prof. Ying TAN

National Laboratory on Machine PerceptionDepartment of Intelligence Science

Peking University, Beijing 100871, P.R.China


Contents• Biological Immune System• Artificial Immune System• Basic Algorithms of AIS • AIS design procedure• Case Studies

– Malicious Executable Detection– Film Recommender

New• Immuneocomputing – IC• Danger Theory• Future


The Immune System is…

Immunity: state or quality of being resistant (immune), either by virtue of previous exposure (adaptive immunity) or as an inherited trait (innate immunity)

Immune system: a system that protects the body from foreign substances and pathogenic organisms by producing the immune response


Why is the Immune System?Immune system has following appealing features:• Recognition

– Anomaly detection– Noise tolerance

• Robustness• Feature extraction• Diversity• Reinforcement learning• Memory; • Dynamically changing coverage• Distributed• Multi-layered• Adaptive


Role of Biological Immune System

• Protect our bodies from pathogen and viruses

• Primary immune response– Launch a response to invading pathogens

• Secondary immune response– Remember past encounters– Faster response the second time around


Immune cells• There are two primarily types of

lymphocytes:– B-lymphocytes (B cells)– T-lymphocytes (T cells)

• Others types include macrophages, phagocytic cells, cytokines, etc.


Where is it?

L y m p h a tic v e s s e ls

L y m p h n o d e s

T h y m u s

S p le e n

T o n s ils a n d a d e n o id s

B o n e m a rro w

A p p e n d ix

P e y e r ’s p a tc h e s

P r im a r y ly m p h o id o r g a n s S e c o n d a r y ly m p h o id o r g a


Multiple layers of the immune system

Phagocyte

Adaptive immune

response

Lymphocytes

Innate immune

response

Biochemical barriers

Skin

Pathogens


Antigen

• Substances capable of starting a specific immune response commonly are referred to as antigens

• This includes some pathogens such as viruses, bacteria, fungi etc .


Biological Immune System

T Cell (Helper)

vsInnate Acquired

vsCell Mediated Humoral

B CellSecretes

AntibodyT Cell (Killer)


How does IS work: A simplistic view

A P C

M H C p r o t e i n A n t i g e n

P e p t i d e

T - c e l l

A c t i v a t e d T - c e l l

B - c e l l

L y m p h o k i n e s

A c t i v a t e d B - c e l l( p l a s m a c e l l )

( I )

( I I I )

( I V )

( V )

( V I )

( V I I )

( I I )


Self/Non-Self Recognition

• Immune system needs to be able to differentiate between self and non-self cells

• Antigenic encounters may result in cell death, therefore– Some kind of positive selection– Some element of negative selection


Immune Pattern Recognition

• The immune recognition is based on the complementaritybetween the binding region of the receptor and a portion of the antigen called epitope.

• Antibodies present a single type of receptor, antigens might present several epitopes.– This means that each antibody can recognize a single

antigen

B-cell

BCR or Antibody

Epitopes

B-cell Receptors (Ab)

Antigen


Clonal Selection

Foreign antigens

Proliferation(Cloning)

Differentiation

Plasma cells

Memory cellsSelection

M

M

Antibody

Self-antigen

Self-antigen

Clonal deletion(negative selection)

Clonal deletion(negative selection)


Main Properties of ClonalSelection (Burnet, 1978)

• Elimination of self antigens• Proliferation and differentiation on contact of

mature lymphocytes with antigen• Restriction of one pattern to one differentiated

cell and retention of that pattern by clonaldescendants;

• Generation of new random genetic changes, subsequently expressed as diverse antibody patterns by a form of accelerated somatic mutation


Immune Network Theory

• Idiotypic network (Jerne, 1974)

• B cells co-stimulate each other– Treat each other a bit like antigens

• Creates an immunological memory

12

3

A g

ActivationPositive response

SuppressionNegative response

Antibody

Paratope

Idiotope


Reinforcement Learning and Immune Memory

• Repeated exposure to an antigen throughout a lifetime

• Primary, secondary immune responses• Remembers encounters

– No need to start from scratch– Memory cells

• Continuous learning


Learning (2)

Antigen Ag1 Antigens Ag1, Ag2

Primary Response Secondary Response

Lag

Response to Ag1

Ant

ibod

y C

once

ntra

tion

Time

Lag

Response to Ag2

Response to Ag1

...

...

Cross-Reactive Response

...

...

Antigen Ag1 + Ag3

Response to Ag1

’=Ag1 + Ag3

Lag


Immune System: Summary

• Define host (body cells) from external entities.• When an entity is recognized as foreign (or

dangerous)- activate several defense mechanisms leading to its destruction (or neutralization).

• Subsequent exposure to similar entity results in rapid immune response.

• Overall behavior of the immune system is an emergent property of many local interactions.

Back


Immune metaphors

Immune System

Idea! Idea ‘

Other areas

Artificial ImmuneSystems

Back


What is an Artificial Immune System?

Dasgupta’99: “Artificial immune systems (AIS) are intelligent and adaptive systems inspired by the immune system toward real-world problem solving”

de Castro and Timmis: “Artificial Immune Systems (AIS) are adaptive systems, inspired by theoretical immunology and observed immune functions, principles and models, which are applied to problem solving”http://www.cs.kent.ac.uk/people/staff/jt6/aisbook/

Definition

•Using natural immune system as a metaphor for solving complex computational problems. •Not modelling the immune system


AI models and their corresponding natural prototypes

Genetic Algorithms (GA)

MolecularGenetic code

Artificial immune systems (AIS)

MolecularMolecules of proteins

Cellular automata (CA)CellsBiological cells

Neural computing (NC) Neural networks (NN)

CellsBrain nervous net

Formal logicFormal linguistic

Left hemisphere of brain

Natural language

AI modelBiological levelNatural prototype


Some History• Developed from the field of theoretical

immunology in the mid 1980’s.– Suggested we ‘might look’ at the IS

• 1990 – Bersini first use of immune algorithms to solve problems

• Forrest et al – Computer Security mid 1990’s

• Hunt et al, mid 1990’s – Machine learning• More……


AIS’ Scope• Pattern recognition;• Fault and anomaly detection;• Data analysis;• Data mining (classification/clustering)• Agent-based systems;• Scheduling;• Machine-learning;• Autonomous navigation and control;• Search and optimization methods;• Artificial life;• Security of information systems; • Optimization;• Just to name a few.


Typical Applications of AIS• Computer Security(Forrest’94’96’98, Kephart’94, Lamont’98’01,02,

Dasgupta’99’01, Bentley’00’01,02)• Anomaly Detection (Dasgupta’96’01’02)• Fault Diagnosis (Ishida’92’93, Ishiguro’94)• Data Mining & Retrieval (Hunt’95’96, Timmis’99’01, ’02)• Pattern Recognition (Forrest’93, Gibert’94, de Castro ’02)• Adaptive Control (Bersini’91)• Job shop Scheduling (Hart’98, ’01, ’02)• Chemical Pattern Recognition (Dasgupta’99)• Robotics (Ishiguro’96’97,Singh’01)• Optimization (DeCastro’99,Endo’98, de Castro ’02)• Web Mining (Nasaroui’02,Secker’05)• Fault Tolerance (Tyrrell, ’01, ’02, Timmis ’02) • Autonomous Systems (Varela’92,Ishiguro’96)• Engineering Design Optimization (Hajela’96 ’98, Nunes’00)

Back


Basic Immune Models and Algorithms

• Bone Marrow Models• Negative Selection Algorithms• Clonal Selection Algorithm• Immune Network Models• Somatic Hypermutation


Bone Marrow Models• Gene libraries are used to create antibodies

from the bone marrow• Antibody production through a random

concatenation from gene libraries• Simple or complex libraries

An individual genome corresponds to four libraries:

Library 1 Library 2 Library 3 Library 4

A1 A2 A3 A4 A5 A6 A7 A8

A3 D5C8B2

A3 D5C8B2

A3 B2 C8 D5

= four 16 bit segments

= a 64 bit chain

Expressed Ab molecule

B1 B2 B3 B4 B5 B6 B7 B8 C1 C2 C3 C4 C5 C6 C7 C8 D1 D2 D3 D4 D5 D6 D7 D8


Negative Selection (NS) Algorithms• Forrest 1994: Idea taken from the negative

selection of T-cells in the thymus• Applied initially to computer security• Split into two parts:

– Censoring– Monitoring

Selfstrings (S)

Generaterandom strings

(R0)Match Detector

Set (R)

Reject

No

Yes

N o

Y es

D etec to r S e t(R )

P ro tectedS trings (S ) M a tch

N on-se lfD e tec ted

CensoringMonitoring


Clonal Selection Algorithm (de Castro & von Zuben, 2001)

1. Initialisation: Randomly initialise a population (P)2. Antigenic Presentation: for each pattern in Ag, do:

2.1 Antigenic binding: determine affinity to each P2.2 Affinity maturation: select n highest affinity from P and clone and mutate prop. to affinity with Ag, then add new mutants to P

3. Metadynamics: 3.1 select highest affinity P to form part of M3.2 replace n number of random new ones

4. Cycle: repeat 2 and 3 until stopping criteria (e.g. Max Generation)


CLONALG for PR, Learning, Optimization

Ab {r}

Ab {m}

fj

Agj

Select

Ab {n}

Clone

CjSelect

Cj*

Fj*

Select

Abj*

Ab{d}

L.N. de Castro, et.al., Learning and optimization using the clonal selection principle, IEEE Trans. Evolutionary computation, vol.6, no.3, June 2002, pp.239-251


Discrete Immune Network Models (Timmis & Neal, 2001)

1. Initialisation: create an initial network from a sub-section of the antigens2. Antigenic presentation: for each antigenic pattern, do:

2.1 Clonal selection and network interactions: for each network cell,determine its stimulation level (based on antigenic and network interaction)2.2 Metadynamics: eliminate network cells with a low stimulation2.3 Clonal Expansion: select the most stimulated network cells andreproduce them proportionally to their stimulation2.4 Somatic hypermutation: mutate each clone 2.5 Network construction: select mutated clones and integrate

3. Cycle: Repeat step 2 until termination condition is met


Immune Network Models• Timmis & Neal, 2000• Used immune network theory as a basis,

proposed the AINE algorithmInitialize AINFor each antigen

Present antigen to each ARB in the AINCalculate ARB stimulation levelAllocate B cells to ARBs, based on stimulation levelRemove weakest ARBs (ones that do not hold any B cells)

If termination condition metexit

elseClone and mutate remaining ARBsIntegrate new ARBs into AIN


Immune Network Models• De Castro & Von Zuben (2000c)• aiNET, based in similar principles

At each iteration step doFor each antigen do

Determine affinity to all network cellsSelect n highest affinity network cellsClone these n selected cells

Increase the affinity of the cells to antigen by reducing the distance between them (greedy search)

Calculate improved affinity of these n cellsRe-select a number of improved cells and place into matrix MRemove cells from M whose affinity is below a set thresholdCalculate cell-cell affinity within the networkRemove cells from network whose affinity is below

a certain thresholdConcatenate original network and M to form new network

Determine whole network inter-cell affinities and remove all those below the set threshold

Replace r% of worst individuals by novel randomly generated onesTest stopping criterion


Somatic Hypermutation• Mutation rate in proportion to affinity• Very controlled mutation in the natural immune

system• Trade-off between the normalized antibody

affinity D* and its mutation rate α,

0 0 . 1 0 . 2 0 . 3 0 . 4 0 . 5 0 . 6 0 . 7 0 . 8 0 . 9 10

0 . 1

0 . 2

0 . 3

0 . 4

0 . 5

0 . 6

0 . 7

0 . 8

0 . 9

1

D *

α

ρ = 5

ρ = 1 0

ρ = 2 0

Back


General Framework of AIS

Immune Algorithms

Affinity Measures

Representation

Solution

Problem Application Domain


Representation – Shape Space

• Describe the general shape of a molecule

A n t i b o d y

A n t i g e n

•Describe interactions between molecules

•Degree of binding between molecules


Representation

• Vectors Ab = ⟨Ab1, Ab2, ..., AbL⟩Ag = ⟨Ag1, Ag2, ..., AgL⟩

• Real-valued shape-space• Integer shape-space• Binary shape-space• Symbolic shape-space


Define their Interaction• Define the term Affinity• Affinity is related to distance

– Euclidian∑=

−=L

iii AgAbD

1

2)(

• Other distance measures such as Hamming, Manhattan etc. etc.

• Affinity Threshold


Shape Space Formalism

• Repertoire of the immune system is complete (Perelson, 1989)

• Extensive regions of complementarity

• Some threshold of recognition

εVε

εVε

εVε

V

´ ´

´

´

´

´

´


AIS Design• Problem description• Deciding the immune principles used for

problem solving• Engineering the AIS

– Defining the types of immune components used– Defining the representation for the elements of the AIS– Applying immune principle to problem solving– The meta-dynamics of an AIS

• Reverse mapping from AIS to the real problem

Back


Case Studies of AIS• Malicious Executables Detection ---

From Z.H. Guo, Z.K. Liu, and Y. Tan, An NN-based Malicious Executables Detection Algorithm based on Immune Principles, F.Yin, J.Wang, C. Guo (Eds.): ISNN 2004, Springer, Lecture Notes in Computer Science 3174, pp. 675-680, 2004. (http://dblp.uni-trier.de)

• Film Recommender --- From Dr. Dr UweAickelin (http://www.aickelin.com), University of Nottingham, U.K. 2004

Back


Immuneocomputing -- ICBy Tarakanov, A. 2001.Aims of• A proper mathematical framework;• A new kind of computing;• A new kind of hardware.

New concepts offormal protein (FP) ------- vs. neuronformal immune networks (FIN)------- vs. NN

New!

•A.O. Tarakanov, V.A. skormin, and S.P. Sokolova, Immunocomputing: Principles and Applications, Springer, 2003.Refer to


Problems of Traditional Self/Non-self View

• No reaction to foreign bacteria in gut (friendly bacteria…).

• No reaction to food / air / etc.• The human body changes over its life.• Auto-immune diseases.• How do we produce antibodies that react against

antigens and yet avoid self?• Is it necessary to attack all non-self or a specific self?


The Danger Theory• In the danger model, the idea is to recognise ‘danger’

rather than non self.

• The screening is accomplished post production through an external ‘danger’ signal. Thus the production of autoreactive antibodies (which react to self) is allowed.

• If an (e.g. autoreactive) antibody matches a stimulus in the absence of danger, it is removed. Thus harmless antigens are tolerated, and changing self accommodated.

Matzinger (2002). The Danger Model: A renewed sense of self , Science 296: 301-304.

New!


Danger Theory (con’t)

• Danger Theory – Not self/non-self but Danger/Non-Danger– Immune response is initiated in the tissues.

Danger Zone.– This makes it context dependant

• Matzinger (2002) The Danger Model: A renewed sense of self Science 296: 301-304• Aickelin & Cayzer (2002) The Danger Theory and Its Application to Artificial Immune Systems, Proc. International Conference on AIS (ICARIS 2002)


Danger Zone

Antigens

Antibodies

Match, buttoo faraway

Stimulation

DangerZone

Danger Signal

Damaged Cell

Cells

No match


Towards a ‘dangerous’ IDS

Aickelin U, Bentley P, Cayzer S, Kim J and McLeod J (2003): 'Danger Theory: The Link between AIS and IDS?', Proceedings ICARIS-2003, 2nd International Conference on Artificial Immune Systems, LNCS 2787, pp 147-155

“The danger theory suggests that the immune system reacts to threats based on the correlation of various (danger) signals, providing a method of ‘grounding’ the immune response, i.e. linking it directly to the attacker.”


Other ways of using dangerDanger = Crime, Antigen = Suspect

or...

Danger = Context ?

It could also be useful for data mining, where the ‘danger’signal is a proxy measure of interest

‘Danger Zone’ can be spatial or temporalAndrew Secker, Alex Freitas, and Jon Timmis (2005) “Towards a danger theory inspired artificial immune system for web mining” in A Scime, editor, Web Mining: applications and techniques, pages 145-168 (Idea Group)


Some Recent Applications of Danger Theory

• Anjum Iqbal, Mohd Aizaini Maarof, “Danger Theory and Intelligent Data Processing,”International Journal of Information Technology, Vol.1, No.1, 2004.

• Andrew Secker, Alex A. Freitas, and Jon Timmis, “A Danger Thory Inspired Approach to Web Mining,” Computing Lab. University of Kent, Canterbury, Kent, UK.2005

• So on.

Back


The Future• More formal approach required?• Wide possible application domains.• What makes the immune system

unique?• More work with immunologists:

– Danger theory.– Idiotypic Networks.– Self-Assertion.


Reference for further readingBooks• Artificial Immune Systems and Their

Applications by Dipankar Dasgupta (Editor) Springer Verlag, January 1999.

• L.N. de Castro and J. Timmis, Artificial Immune Systems: A New Computational Intelligence Approach, Springer, 2002.

• A.O. Tarakanov, V.A. skormin, and S.P. Sokolova, Immunocomputing: Principles and Applications, Springer, 2003.

Related academic papers• J. Timmis, P.Bentley, and Emma Hart (Eds.): Artificial Immune Systems,

Proceedings of Second International Conference, ICARIS 2003, Edinburgh, UK, September 2003. LNCS 2787, Springer.


New Events:• Special Session on Artificial Immune Systems at the Congress

on Evolutionary Computation (CEC), December 8-12, 2003, Canberra, Australia.

• Special Session on Immunity-Based Systems at Seventh International Conference on Knowledge-Based Intelligent Information & Engineering Systems (KES), September 3-5, 2003, University of Oxford, UK.

• Second International Conference on Artificial Immune Systems (ICARIS), September 1-3, 2003, Napier University, Edinburgh, UK.

• Tutorial on Artificial Immune Systems at 1st Multidisciplinary International Conference on Scheduling: Theory and Applications (MISTA), 12 August 2003, The University of Nottingham, UK.

• Tutorial on Immunological Computation at International Joint Conference on Artificial Intelligence (IJCAI), August 10, 2003, Acapulco, Mexico.

• Special Track on Artificial Immune Systems at Genetic and Evolutionary Computation Conference (GECCO), Chicago, USA, July 12-16, 2003


AIS Resources• Artificial Immune Systems and Their Applications by D

Dasgupta (Editor), Springer Verlag, 1999.• Artificial Immune Systems: A New Computational

Intelligence Approach by L de Castro, J Timmis, Springer Verlag, 2002.

• Immunocomputing: Principles and Applications by A Tarakanov et al, Springer Verlag, 2003.

• Third International Conference on Artificial Immune Systems (ICARIS), September 13-16, 2004, University of Catania, Italy.

• 4th International Conference on Artificial Immune Systems(ICARIS), 14th-17th August, 2005 in Banff, Alberta, Canada


That’s all

First Page


Malicious Executables Detectionbased on Artificial Immune Principles*

* This work was supported by Natural Science Foundation of China with Grant No. 60273100.

From Z.H. Guo, Z.K. Liu, and Y. Tan, An NN-based Malicious Executables Detection Algorithm based on Immune Principles, F.Yin, J.Wang, C. Guo (Eds.): ISNN 2004, Springer Lecture Notes on Computer Science 3174, pp. 675-680, 2004. (http://dblp.uni-trier.de)

Case Study 1:


Outline• Definition of Terms• Goal and Motivation• Previous Research works• Immune Principle for Malicious Executable

Detection• Malicious Executable Detection Algorithm • Experiments and Discussion• Concluding Remarks


Definition of Terms• Malicious Executable

is generally defined as a program that has some malicious functions, such as compromising a system’s security, damaging a system or obtaining sensitive information without the permission of users. It includes virus, trojanhorse, worm etc.

• Benign Executableis a normal program without any malicious function.

Back


Dos/Win32 viruses

Computers / Information SystemsTrojan horses

Worms

eMail attached viruses

Malicious executables

tens of thousands of new viruses / year

Appear!

But: Current antivirus systems attempt to detect these new malicious programs with heuristics by hand (costly and ineffective)

Current Task:Devise new methods for detecting new ME


Definition of Symbols and Structures

B: binary code alphabet, B={0,1}.Seq(s,k,l): short sequence cutting operation. Supposing s is binary sequence, and s＝b(0)b(1)…b(n-1), b(i)∈B, then Seq(s,k,l)=b(k)b(k+1)…b(k+l-1).E(k): executable set, k∈{m,b}，m denotes malicious executable, b benign executable.E: whole set of executables, i.e., E= E(m)∪E(b).e(fj,n): executable as binary sequence of length n, and fj is executable identifier.ld: detector code length.lstep: step size of detector generation.dl: detector, dl = Seq(s,k,l).Dl: set of detector with code length l, i.e., Dl ={ dl (0), dl (1),…, dl (nd-1)}, |Dl|= nd.

Back


Goal and Motivation

• Aiming at developing an automatic detection approach of new malicious executables.

• Aiming at trying to use artificial immune system (AIS) and artificial neural networks (ANN), to detect malicious executable with a high Detection Rate (DR) with low False Positive Rate (FPR) over others.

Back


Previous Related Works

• Signature-based Methods

• Expert Knowledge-based Methods

• Machine Learning Methods

Back


Signature-based MethodsIt creates a unique tag for each malicious program so that future

examples of it can be correctly classified with a small error rate. And relies on signatures of known malicious executable to generate detection models.

Drawbacks:• Can not detect unknown and mutated viruses. • As increase of the number and type of viruses, its detection speed

become slow dramatically. At the same time, the analysis of the signatures of viruses become very difficult, in particular, for the encrypted signatures.

(refer to IBM Anti-virus Group’s report: R.W. Lo, K.N. Levitt, and R.A. Olsson. MCF: a Malicious Code Filter. Computers & Security, 14(6):541–566., 1995.)

Back


Expert Knowledge-based Methods

Using the knowledge of a group of virus experts to construct heuristic classifiers for detection of unknown viruses.

Drawbacks:• Time-consuming analysis method.• Only discover some unknown viruses, but its false

detection rate is very high. For detecting unknown virus based on ANN, IBM Anti-virus

Group also proposes one method to detect Boot Sector viruses only.

(refer to W. Arnold and G. Tesauro. Automatically Generated Win32 Heuristic Virus Detection. Proceedings of the 2000 International Virus Bulletin Conference, 2000.)

Back


Machine Learning Methods

• M.G. Schultz developed a framework that used data mining algorithms, i.e., Multi-Naïve Bayesmethod, to train multiple classifiers on a set of malicious and benign executables to detect new examples (unknown ME).

(refer to M.G. Schultz.,E. Eskin and E. Zadok . Data Mining Methods for Detection of New Malicious Executables. IEEE Symposium on Security and Privacy, May 2001.)

Back


Biologically-motivated Information Processing Systems

• Brain-nervous systems – Neural Networks (NN)• Genetic systems – Genetic Algorithms(GA)• Immune systems – Artificial Immune Systems(AIS)

or immunological computation.

NN and GA have extensively studied with wide applications but AIS has relative few applications


Natural prototypes vs. their models

Genetic Algorithms (GA)

MolecularGenetic code

Artificial immune systems (AIS)

MolecularMolecules of proteins

Cellular automata (CA)

CellsBiological cells

Artificial Neural networks (ANN)

CellsBrain nervous net

Formal logicFormal linguistic

Left hemisphere of brain

Natural language

Computing modelBiological level

Natural prototype


Comparison of Three Algorithms

Recognition / Objective FunctionExternal StimuliFitness FunctionInteraction with

Environment

Recognition / Network ConnectionsNetwork ConnectionsCrossoverInteraction between

Components

Recruitment / Elimination of Components

Construction / Pruning of Connections

Recruitment / Elimination of ComponentsMeta-Dynamics

Evolution / LearningLearningEvolutionDynamics

Component Concentration / Network

ConnectionsConnection StrengthsChromosome StringsKnowledge Storage

Discrete components / Networked ComponentsNetworked ComponentsDiscrete ComponentsStructure

DynamicPre-DefinedDynamicLocation of Components

Attribute StringsArtificial NeuronsChromosome StringsComponents

AISNN (Classification)GA (Optimisation)


Immune Principles for Malicious Executable Detection

• Non-self Detection Principle

• Anomaly Detection Based on Thickness

• The Diversity of Detector Representation vs. Anomaly Detection Hole

Back


Non-self Detection Principle• For natural immune system, all cells of body are

categorized as two types of self and non-self. The immune process is to detect non-self from cells.

• To realize the non-self detection, the maturation process of lymphocytes T cell undergoes two selection stages of Positive Selection and Negative Selection since antigenic encounters may result in cell death. Some computer scientists inspired by these two stages had proposed some algorithms used to detect anomaly information. Here, we will use the Positive Selection Algorithm (PSA) to perform the non-self detection for recognizing the malicious executable.


Non-self Detection by PSA

Match ?？

Y

N

Detector Set Dl

self non-self

Short sequence to be detected

（Its length is l）

Process of anomaly detection with PSA

Back


Anomaly Detection Based on Thickness

• Anomaly recognition process is one process that immune cells detect antigens and are activated.

• The activated threshold of immune cells is decided by the thickness of immune cells matching antigens.

Back


The Diversity of Detector Representation vs. Anomaly Detection Hole

• The main difficulty of anomaly detection is utmost decreasing the anomaly detection hole. The natural immune system resolves this problem well by use of the diversity of MHC (Major Histocompatibility Complex) cell representations, which decides the diversity of anti-body touched in surface of T cells. This property is very useful in increasing the power of detecting mutated antigens, and decreasing the anomaly detection hole.

• According to the principle, we can use the diversity of detector representation to decrease the anomaly detection hole. As was illustrated by following schematic drawings.


Schematic diagram of abnormal detection holes (cont’)

Self Space

Nonself Space

Detectors

Abnormal detection holes


Reduction of abnormal detection holes by use of the diversity of

detector representations

Detector Representation 1



Combination of detectors

back


Malicious Executable Detection Algorithm (MEDA)

MEDA based on AIS includes three parts,

• Detector generation,

• Anomaly information extraction ,

• and Classification.


Flow Chart of Malicious Executable Detection Algorithm (MEDA)

Executable to be detected(…00111101…) Output

Update Gene(…10101101…)

Generating detector set

Extracting anomaly property

Classifier

Gene(…01101001…)

MEDA

Back


Generation of Detector SetDetector generation algorithm:

• Begin initialize lstep、ld、k=0• Do cutting e(fk,n) from Eg(b)• i=0；• While i <= n-ld-1 do • Begin• d = Seq(e(fk,n),i, ld)；• if d≠€ Dld then Dld←d；• i=i+lstep；• End• k=k+1；• Until Eg(b) is empty• Return Dld ；• End


Illustration of Detector Generating Process

File Hex Sequence: 56 32 12 0A 34 ED FF 00 2D…. . 00 0A 34 ED FF FA 11 00Extracting Detector: 56 32 12

32 12 0A12 0A 34

┋……………………………………………┋

FF FA 11FA 11 00

Generating Process of 24-bit Detectors with 8-bit stepsize (ld=24, lstep=8)

Back


Extraction of Anomaly Characteristics --Non-self Thickness (NST)

• Non-self Detection• NST, as Anomaly Property, is defined

as the ratio of number of non-self units to file binary sequence, pl=nn/(nn+ns).

• If there are m kinds of detectors, the file has a NST Vector P=(pl1, pl2, …, plm)T.


NST Extraction Diagram

“Nonself” Detection

File to be detected（…00111101）

Is “Nonself” ?

ns add 1 nn add 1

Y

N

Completing detection ？Y

N

Compute pl=nn/(nn+ns)

Initialization，choose lstep、ld , Dl

End


NST Extraction Algorithm• Begin open e(fk,n);• Select lstep, ld; • Set ns=0, nn=0, i=0;• While i <= n-ld-1 do • Begin• s = Seq(e(fk,n),i, ld)；• if s ≠€ Dld then nn = nn+1；• else ns = ns + 1；• i = i + lstep；• End• pld = nn / ( ns＋nn )；• Return pld ；• End

Back


BP Network Classifier• We use Anomaly Property Vector

(APV), i.e., NST vector P, as input variable of two-layer BP network classifier. The number of nodes of input layer equals to APV’s dimension.

• The Sigmoid transfer function is chosen for the hidden layer and Linear function for the output layer.


BP Network Classifier Structure

pl1

pl2

plm

Out (1-ME, 0-BE)P

Non-S

elf Thickness (NS

T) Vector

Back


Experiments and Discussion• Experimental Data Set

• Generation of Detector Set

• Experimental Result Using Single Detector Set

• Experimental Result Using Multi-Detector Set

Back


Experimental Data Set

• BE—Benign Executable• ME—Malicious Executable

All Justified by Antivirus cleaner Tools

4481Total

DOS virus, Win32 virus, Trojan, Worm, etc. from Internet.

3566 ME

Win 2K OS and some application programs.

915BE

RemarksFilesType

Back


Generation of Detector Set

• Eg(b) is Gene of generating detector, ld ∈{16，24，32，64，96}, and lstep=8bits. By using the detector generating algorithm, we can get D16, D24, D32, D64, and D96, separately.

Table1: Detectors generation result

TreeTreeTreeBitmap Index

Bitmap Index

storestructure

21,294,857

12,768,361

8,938,352

10,931,627

65536|Dld|

9664322416Code Length ld

Back


Detection Result of Malicious Executables by D24

异己

”浓

度P2

4

(a) NST of files, where symbol ‘x’ represents benign program (Red),

‘□’ malicious program (Blue)

正确检测

率(Detection Rate)%

误判率(False Positive Rate)%

(b) ROC Curve

File No.

NS

T p24



文件序号

异己

”浓

度P32

正确

检测

率(Detection Rate)%


(b) ROC Curve(a) NST of files, where symbol ‘x’ represents benign program,

‘□’ malicious program

NS

T p32



文件序号

异己

”浓

度P64

正确

检测率

(Detection Rate)%


(b) ROC Curve(a) NST of files, where symbol

‘x’ represents benign program (Red), ‘□’ malicious program (Blue)

NS

T p64


Experimental Result Using Single Detector Set

0 2 0 4 0 6 0 8 0 1 0 00

2 0

4 0

6 0

8 0

1 0 0

1 6 b its D a ta S e t2 4 b its D a ta S e t3 2 b its D a ta S e t6 4 b its D a ta S e t9 6 b its D a ta S e t

F a l s e P o s i t i v e R a t e ( % )

Det

ectio

n R

ate

(%)


When FPR is fixed, relationship curves of DR versus Code

Length ldDetection Rate (%)

C o d e le n g th l d（ b its）

Note: from the bottom to up, the FPR is 0%, 0.5%, 1%, 2%, 4%, 8%, and 16%, in sequence.

Back


Experimental Result Using Multi-Detector Set

• This experiment selects multi-detector set to detect benign and malicious executables.

• We don’t use D16 because of its zero DR and also set D96 as upper limit because almost same DR values when ld ≥96.

• Here we selects D24, D32, D64 and D96 four detector sets as anomaly detection data set, and uses them to extract Non-self thickness (NST) vector, and finally a BP network is exploited as classifier.

• For the process of classification, we randomly selects 30% files of E(b) as Eg(b) to train a BP network, and use the remaining data to illustrate the anomaly detection performance.


NST Distribution and ROC Curve of Multi-Detector Set Method

Detection Rate (%)

False Positive Rate (%)

“异

己”

浓度

（64bits）

“异己”

浓度（24bits）

“异己”浓度（32bits）

(a) NST of files for mixture of D24, D32 and D64.

‘x’ benign program (in Red),‘□’ malicious program (in Blue).

(b) ROC Curve of mixed detector set of D24, D32, D64 and D96


Comparisons With Bayes Methods and Signature-based Method

0 2 4 6 8 1 0 1 20

2 0

4 0

6 0

8 0

1 0 0

M E D A w ith B P N e tw o r kN a ive B a ye s w ith S tr in g sM u lti- N a ive B a ye s w ith B yte sS ig n a tu r e M e th o d

F a l s e P o s i t i v e R a t e ( % )

Det

ectio

n R

ate

(%)


Algorithm Complexities

ComputingJoint Probs.

Computing NST

Name

Operation type 3

1Gblf float multiplica-tions

Depend on P(Fi/C)

Searching P(Fi/C)

>>ltrainProb. Info.

Bayes

0.4Gb4×lf additions

≤80×ltest

detector matching

ltraindetectorsMEDA

AmountAmountNameAmountNameAlgorithm

Store Space

Operation type 2Operation type 1

1

( ) ( / )n

ii

P C P F C=∏

Back


Remarks• For short binary sequence and single detector

set for the detection of malicious executables, the performance of D24 is the best, giving out DR 80.6% with FPR 3%.

• For long code length of detector and multi-detector set, our method obtains the best performance of DR 97.46% with FPR 2%, over current methods.

• This result verifies – diversity of detector representation can decrease

anomaly detection holes. – “non-self” thickness detection.

Back

Back


Film Recommender

Prediction:

– What rating would I give a specific film?

Recommendation:

– Give me a ‘top 10’ list of films I might like.

From Dr. Dr Uwe Aickelin (http://www.aickelin.com)University of Nottingham, U.K.,

Case Study 2:


Film Recommender (con’t 1)

• EachMovie database (70k users).• User Profile: set of tuples {movie, rating}.• Me: My user profile.• Neighbour: User profile of others. • Similarity metric: Correlation score.• Neighbourhood: Group of similar users.• Recommendations: From neighbourhood.



• User Profile: set of tuples {movie, rating}• Me: My user profile.• Neighbour: User profile of others. • Affinity metric: Correlation score.

• Neighbourhood: Group of similar users.

• Recommendations: From neighbourhood

Antigen

Antibody

Antibody – Antigen Binding Antibody – Antibody Binding

Group of antibodies similar to antigen and dissimilar to other antibodies

Stim

ulat

ion

Suppression

Weighted Score based on Similarities.



• Start with empty AIS.• Encode target user as an antigen Ag.• WHILE (AIS not full) && (More Users):

– Add next user as antibody Ab.– IF (AIS at full size) Iterate AIS.

• Generate recommendations from AIS.


Film Recommender (con’t 4)Suppose we have 5 users and 4 movies:

– u1={(m1,v11),(m2,v12),(m3,v13)}.– u2={(m1,v21),(m2,v22),(m3,v23),(m4,v24)}.– u3={(m1,v31),(m2,v32),(m4,v34)}.– u4={(m1,v41),(m4,v44)}.– u5={(m1,v51),(m2,v52),(m3,v53), (m4,v54)}.

• We do not have users’ votes for every film.

• We want to predict the vote of user u4 on movie m3.


Algorithm walkthrough (1)

Start with empty AIS:DATABASE

u1, u2, u3, u4, u5

AIS

User for whom to predict becomes antigen:

DATABASE

u1, u2, u3, u5

u4

AIS

Ag



Add antibodies until AIS is full…

Ab1

DATABASE

u2, u3, u5

u1

AIS

Ag

DATABASE

u4

u2,u3

AIS

Ag

Ab1 Ab2

Ab3



• Table of Correlation between Aband Ag:– MS14, MS24, MS34.

• Table of Correlation between Antibodies:– MS12 = CorrelCoef(Ab1, Ab2) – MS13 = CorrelCoef(Ab1, Ab3)– MS23 = CorrelCoef(Ab2, Ab3)

Ab2

Ab3

Ab1

Ag



• Calculate Concentration of each Ab:– Interaction with Ag (Stimulation).– Interaction with other Ab (Suppression).

Ab1

Ab1 Ab2

Ab2

Ab2

Ab2Ab2

Ag

Ab1 Ab2

Ab3

AIS

Ag

AIS



• Generate Recommendation based on Antibody Concentration.

Recommendation for user u4 on movie m3will be highly based on vote on m3 of user u2

AIS

Ab1

Ab2

Ab1

AgAb2

Ab2

Ab2Ab2


Film Recommender Results

• Tested against standard method (Pearson k-nearest neighbours).

• Prediction:– Results of same quality.

• Recommendation:– 4 out of 5 films correct (AIS).– 3 out of 5 films correct (Pearson).

Back

Documents

Artificial Immune Systems AIS Model Tutorial