Adversarial Information Retrieval The Manipulation of Web Content

Introduction Examples TrustRank and Other Methods

What is Adversarial IR? Gathering, Indexing, Retrieving and Ranking Information Subset of the information has been manipulated maliciously Financial Gain

What is the Goal of AIR? Detect the bad sites or communities Improve precision on search engines by eliminating the bad guys

Simplest form First generation engines relied heavily on tf/idf – The top-ranked pages for the query maui resort were the ones containing the most maui’ s and resort’ s SEOs responded with dense repetitions of chosen terms – e.g., maui resort maui resort maui resort – Often, the repetitions would be in the same color as the background of the web page Repeated terms got indexed by crawlers But not visible to humans on browsers Pure word density cannot be trusted as an IR signal

Search Engine Spamming Link-spam Link-bombing Spam Blogs Comment Spam Keyword Spam Malicious Tagging

Spamming Online tutorials for “search engine persuasion techniques” – “How to boost your PageRank” Artificial links and Web communities Latest trend: “Google bombing” – a community of people create (genuine) links with a specific anchor text towards a specific page. Usually to make a political point

Google Bombing

Our Focus Link Manipulation

Trust Rank Observation – Good pages tend to link good pages. – Human is the best spam detector Algorithm – Select a small subset of pages and let a human classify them – Propagate goodness of pages 10

Propagation Trust function T – T(p) returns the propability that p is a good page Initial values – T(p) = 1, if p was found to be a good page – T(p) = 0, if p was found to be a spam page Iterations: – propagate Trust following out-links – only a fixed number of iteration M. 11

Propagation (2) Problem with propagation –Pages reachable from good seeds might not be good –the further away we are from good seed pages, the less certain we are that a page is good. 12 – solution: reduce trust as we move further away from the good seed pages (trust attenuation).

Trust attenuation – dampening –Propagate a dampened trust score ß < 1 at first step –At n-th step propagate a trust of ß^n 13

Trust attenuation – splitting –Parent trust value is splittet among child nodes –Observation: the more the links the less the care in choosing them –Mix damp and split? ß^n(splitted trust) 14

Selection – Inverse PageRank The seed set S should: – be as small as possible – cover a large part of the Web Covering is related to out-links in the very same way PageRank is related to in-link – Inverse PageRank ! Perform PageRank on a graph with inverted links –G' = (V, E') where ( p, q )  E'  ( q, p )  E. 15

Algorithm 1.Select seeds ( s ) and order by preference 2.Invoke oracle (human) on the first L seeds, 3.Initialize and normalize oracle response d 4.Compute TrustRank score (as in PageRank formula): t * = ß · T · t *+( 1 − ß ) · d T is the adjacency matrix of the Web Graph. ß is the dampening factor. (usually.85) 16

Algorithm - example –s = [ 0. 08, 0. 13, 0. 08, 0. 10, 0. 09, 0. 06, 0. 02] –Ordering = [ 2, 4, 5, 1, 3, 6, 7 ] –L=3 {2, 4, 5} d =[0, 0.5, 0, 0.5, 0, 0, 0] –ß =0.85 M=20 –t * = [ 0, 0. 18, 0. 12, 0. 15, 0. 13, 0. 05, ] –NB. max=0.18 –Issues with page 1 and 5 17

Issues with TrustRank Coverage of the seed set may not be broad enough  Many different topics exist, each with good pages TrustRank has a bias towards communities that are heavily represented in the seed set  inadvertently helps spammers that fool these communities

Bias towards larger partitions Divide the seed set into n partitions, each has m i nodes t i : TrustRank score calculated by using partition i as the seed set t : TrustRank score calculated by using all the partitions as one combined seed set

Basic ideas Use pages labeled with topics as seed pages  Pages listed in highly regarded topic directories Trust should be propagated by topics  link between two pages is usually created in a topic specific context

Topical TrustRank  Partition the seed set into topically coherent groups  TrustRank is calculated for each topic  Final ranking is generated by a combination of these topic specific trust scores Note  TrustRank is essentially biased PageRank  Topical TrustRank is fundamentally the same as Topic- Sensitive PageRank, but for demoting spam

Combination of trust scores Simple summation  default mechanism just seen Quality bias  Each topic weighted by a bias factor  Summation of these weighted topic scores  One possible bias: Average PageRank value of the seed pages of the topic

Further Improvements Seed Weighting  Instead of assigning an equal weight to each seed page, assign a weight proportional to its quality / importance Seed Filtering  Filtering out low quality pages that may exist in topic directories Finer topics  Lower layers of the topic directory