4 Adversarial Activity

The more spam a user gets, the less likely he will be to notice one innocent mail sitting in his spam folder. And strangely enough, the better your spam filters get, the more dangerous false positives become, because when the filters are highly effective, users will be more likely to ignore everything they catch. The methods of spamming are improving together with the methods of spam filtering. Spammers try to attack filters, namely to decrease filtering effectiveness. We can categorize several common attacks on spam filters as following:

• Tokenization attacks, when the spammer intends to prevent correct tokenization of the message by splitting or modifying features, for example putting extra spaces, or symbols like ’-’ or ’#’, in the middle of the words, ’U_N_$$_U_B_S_C_R_l_B_E’, for instance.
• Text Obfuscation attacks, when the content of the message is obscured from the filter.
  • deliberately misspelling certain key words, such as ’Wildcatt’, ’weeek’, ’sunsit’, ’veeery’.
  • transposing letters or by substituting letters with lookalike numbers and symbols, for example, ’drgs’ or ’UN$UBB$CRlBE’.
  • miss used accented characters are common to appear.

4.1 Remedies

• Text De-obfuscation
  Spammers often substitute characters (sometimes called LEET) for other characters that look alike in order to avoid known spamming keywords. For example, if you have a content filter that traps messages containing ”viagra”(because of the flood of viagra spasm), all the spammer has to do is spell it as ”v1@gr@”or ”iagra” or even ”viagara” and it will probably sneak past.
• Spell Correction
  In September 2003, a PhD student at Cambridge University had discovered that as long as the first and last letters of a word remained in place , ordering for the others does not matter. For instance, people are able to read the word ”sfotwrae” or ”sofwtare”.

  The original text passed around reads as follows:
  

  Aoccdrnig to rscheearch at Cmabrigde uinervtisy, it deosn’t mttaer waht oredr the ltteers in a wrod are, the olny iprmoetnt tihng is taht thefrist and lsat ltteres are at the rghit pclae. The rset can be a tatol mses and you can sitll raed it wouthit a porbelm. Tihs is bcuseae we donot raed ervey lteter by it slef but the wrod as a wlohe.

  The actual origin of the theory appears to be an unpublished PhD thesis [9].

  After the technique became known on the internet, some people have written simple scripts ⁴ that can randomize text excluding the first and last letters.

  Nowadays, spammers are constantly searching for new means to obfuscate a word in order to fool spam filters, and this technique often appeared in spam text:
  

  Do you wnat to l00k c00l and w3althy but do not have the m0ney to aff0rd a= sweeeeet n3w R0lex wtach? Get a 98% L00kalike R0lex watc_h here! We have replika_s of all the fines_t bran_d watc_hes. Check the_m out here!

• Edit Distance
  The Levenshtein distance is a metric for measuring the amount of difference between two sequences (of characters). The difference between two strings is given by the minimum number of operations needed to transform one string into the other, where an operation is an insertion, deletion, or substitution of a single character.
  

  Levenshtein-Distance(chars[1..m],chart[1..n])
   

  1 Initialize d[0..m, 0..n] = 0. 2 for i ← 0 to m 3 do d[i, 0] = i 4 for j ← 0 to n 5 do d[0,j] = j 6 for i ← 1 to m 7 do for j ← 1 to n 8       do if s[i] = t[j] 9                then cost = 0 10                else cost = 1 11             d[i,j] = min( 12             d[i - 1,j] + 1, // deletion 13             d[i,j - 1] + 1, // insertion 14             d[i - 1,j - 1] + cost) // substitution 15 return d[m,n]

• Probabilistic Approach
  Here are three essential statements that would needs to be addressed:
  • P(w), is the probability of a given words w to appear in an English text.
  • P(c), is the probability that a proposed correction c stands on its own. This is equivalent to the statement ”how likely is c to appear in an English text?” In that sense, P(”the”) would have a relatively high probability, while P(”aaaaabbbbbcccc”) would be near zero.
  • P(w|c), is the probability that w would be typed in a text when the author meant c. That is, ”how likely is it that the author would type w by mistake when c was intended?”

  Given a word w, we are trying to choose the most likely spelling correction for that word . Our goal is to find the correction c, out of all possible corrections, that maximizes the probability of c given the original word w: P(c|w). This is equivalent of saying ”What is the most likely spelling correction c, if user types a word w”. (which could mean itself, if no better correction can be found)

  

  By Bayes’ Theorem this is equivalent to:

  argmax cP(c|w)	= argmax c
  	=∝ argmax cP(w|c)P(c)	(Since P(w) is the same for every possible c, we can ignore it)