The features are the hardest part of this format, as they are the most complex. However, having a complete description of the feature-set used in a paper in this format allows the use of an extractor tool to automatically reproduce the feature vectors used in the paper.


  • Features

    A feature represents a value which can be extracted from a packet/flow/flow aggregation. These can be base features (often can be computed by looking only at packet headers), or some more complicated things (like the entropy of a value that can be found in the packet headers). In our format, each feature is a combination of operations applied to base features. Each feature must be only one scalar value (as opposed to a vector).

  • Base features

    Base features are the basic elements of our features format. These are always represented by strings.

    Examples: "packetTotalCount", "octetTotalCount", "sourceIPv4Address"

  • Operations

    We have multiple operations defined, which receive features as arguments. These are defined as a JSON dictionary with one key, in which the key is the name of the operation, and the corresponding value is a list of arguments.

    Examples: "mean", "add", "log"

  • Selections

    We also have an option to filter out specific packets in a flow/flow aggregation. This allows us for example to count the number of packets with a specific property (e.g., packet size larger than some threshold).

Base Features

We call base features those which are not obtained by combining other features. These are represented in this format by JSON strings.

We try to use the names of the IPFIX information elements defined by IANA. For features that we can not get out of combining IANA features with our limited set of operations, we have two naming options:

  • if the feature is expected to be used many times (e.g.: there are some KDD ‘99 features which we cannot represent using IANA features and operations, but they are used in many papers), use a _ as prefix to a descriptive feature name. This features are listed in own_ies.csv. If you want to specify a new _ feature, you need to add it there.
  • if the feature is very specific to this paper, use __ (double _) as prefix to a descriptive feature name

In both of this cases, try to give descriptive feature names, similar to the the ones used by IANA. Names should use camel case and start with a lower-case character. They should follow the following regex: ^[_]{1,2}[a-z0-9]+([A-Z][a-z0-9]*)*$.

This means that all base features that do not start with _ have to be IPFIX information elements defined by IANA.

There is still another case, which is features that are repeated often, and are a combination of IANA features. In this case, use a descriptive feature name which starts with _ as an alias for it. A complete list of aliases is in feature_aliases.json; please add additional aliases there.


Besides the base features, we also have some operations, by which we can get new features.

We can have two kinds of operations:

  • value
    The output is a single scalar value.
  • values
    The output is a vector of values (possibly of variable size).


The highest level operation in a feature cannot be one that is defined in the <values> directive, as that outputs multiple values.

Below is a grammar defining the list of possible operations, and their respective arguments:


# <value> always outputs a single number (a <value>)
<value> -> {"mean": [<values>]}
<value> -> {"stdev": [<values>]}
<value> -> {"variance": [<values>]}
<value> -> {"median": [<values>]}
<value> -> {"quantile": [<values>, <value>]} # second argument is a number from 0 to 1, where 0 is the minimum and 1 the maximum
<value> -> {"minimum": [<values>]} | {"minimum": [<value>, <value>+]}
<value> -> {"maximum": [<values>]} | {"maximum": [<value>, <value>+]}
<value> -> {"argmin": [<values>]} | {"argmin": [<value>, <value>+]}
<value> -> {"argmax": [<values>]} | {"argmax": [<value>, <value>+]}
<value> -> {"floor": [<value>]}
<value> -> {"ceil": [<value>]}
<value> -> {"mode": [<values>]} # returns the most frequent element in <values>
<value> -> {"mad": [<values>]} # returns the mean absolute deviation of <values>
<value> -> {"moment": [<values>, <value>]} # returns the <value>-th standardized moment of <values>
<value> -> {"count": [<selection>]}  # returns number of selected objects
<value> -> {"length": [<values>]}  # returns number of values (useful to use with quantile_range)
<value> -> {"distinct": [<values>]}  # returns number of distinct values in <values> in the selected objects
<value> -> {"apply": [<value>, <selection>]}  # returns a single feature value for the selection of objects
<value> -> {"add": [<value>, <value>+]} | {"add": [<values>]}
<value> -> {"subtract": [<value>, <value>]}
<value> -> {"multiply": [<value>, <value>+]} | {"multiply": [<values>]}
<value> -> {"divide": [<value>, <value>]}
<value> -> {"pow": [<value>, <value>]}  # raises the first value to the power of the second value (e.g., pow(octetTotalCount, 2) == octetTotalCount^2)
<value> -> {"log": [<value>]}
<value> -> {"exp": [<value>]}
<value> -> {"entropy": [<values>]}
<value> -> {"get": [<value>, <values>]}  # gets the <value>-th element of the second argument; indexing is like in Python
<value> -> {"get_bits": [<value>, <value>]}  # gets the <value>-th bit of the second argument; indexing is like in Python
<value> -> {"slice_bits": [<value>, <value>, <value>]}  # gets third_argument[first_argument : second_argument], considering the third argument as an array of bits, and returns it as a value; indexing is like in Python
<value> -> {"ifelse": [<logic>, <value>, <value>]}  # if the condition is true, return the first argument else the second
<value> -> {"left_shift": [<value>, <value>]}  # shift the bits in the first value left by the second value
<value> -> {"right_shift": [<value>, <value>]}  # shift the bits in the first value right by the second value


# <values> outputs a list of <value>
<values> -> {"map": [<down>, <selection>]}  # returns a feature value for each object in selection
<values> -> {"slice": [<value>, <value>, <values>]}  # gets third_argument[first_argument, second_argument]; indexing is like in Python
<values> -> {"quantile_range": [<values>, <value>, <value>]} # e.g. {"quantile_range": [<values>, 0, 0.25]} returns all values in the first quartile
<values> -> {"flat_map": [<down2>, <selection>]} | {"flat_map": [<down2>, <selection>, <selection>]}  # only applicable for flow-aggregations; just one selection applies same selection for both flows and packets; two selections applies the 1st selection for flows and the second for packets
<values> -> <down>  # features from one level-down (in flows, packet features; in flow-aggregations, flow features)


The selection directive is useful for filtering out packets or any other information which might not be interesting for a particular feature. Intuitively, using selection on a flow will select packets (that is, the result will be the packets that fulfill the conditions in the selection), and in a flow_aggregation will output either flows or packets, depending on the selection used. Because of this distinction, for each selection that outputs packets, there is another selection that outputs flows, and contains "_flows" in its name.

This distinction between outputting flows or packets is necessary, since you can select objects with "octetTotalCount" > 1000, and in this case it’s ambiguous whether you want to select all packets with more than 1000 bytes, or all the flows with more than 1000 bytes. Note that some features only make sense for flows (e.g., "packetTotalCount").

Its syntax is the following:

# <selection> outputs a list of objects (packets, flows or aggregations, depending on what kind of feature is used)
<selection> -> {"select": [<logic-down>]}
<selection> -> {"select_slice": [<value>, <value>]} | {"select_slice": [<value>, <value>, <selection>]}  # selects a slice from the first value to the second value, with Python-like indexing (if a <selection is not provided, default to selecting everything)
<selection> -> "forward" | "backward"  # special cases for selection; select objects in the forward (or backward) direction

The logic directive contains the test to decide what gets or not filtered. Definition of logic:

# <logic> is used for selection, should be evaluated for each object
<logic> -> {"and": [<logic>+]} 
<logic> -> {"or": [<logic>+]}
<logic> -> {"geq": [<value>, <value>]}
<logic> -> {"leq": [<value>, <value>]}
<logic> -> {"less": [<value>, <value>]}
<logic> -> {"greater": [<value>, <value>]}
<logic> -> {"equal": [<value>, <value>]}
<logic> -> true | false
<logic-down> -> {"and": [<logic-down>+]} 
<logic-down> -> {"or": [<logic-down>+]}
<logic-down> -> {"geq": [<down>, <value>]}
<logic-down> -> {"leq": [<down>, <value>]}
<logic-down> -> {"less": [<down>, <value>]}
<logic-down> -> {"greater": [<down>, <value>]}
<logic-down> -> {"equal": [<down>, <value>]}
<logic-down> -> true | false

Example Features

The following are examples of the features directive.

"features": [
  {"divide": ["octetTotalCount", "_activeForSeconds"]},
  {"divide": ["packetTotalCount", "_activeForSeconds"]},
  {"maximum": ["_interPacketTimeMicroseconds"]},
  {"minimum": ["_interPacketTimeMicroseconds"]},
  {"count": [{"select": [{"geq": ["_interPacketTimeMicroseconds", 1000000]}]}]}
"features": [
  {"entropy": ["sourceIPv4Address"]},
  {"entropy": ["destinationIPv4Address"]},
  {"entropy": ["destinationTransportPort"]},
  {"entropy": ["_flowDurationSeconds"]},
  {"multiply": [{"argmax": [{"count": [{"select": [{"less": ["ipTotalLength", 128]}]}]}, {"count": [{"and": [{"select": [{"geq": ["ipTotalLength", 128]}]}, {"select": [{"less": ["ipTotalLength", 256]}]}]}]}, {"count": [{"and": [{"select": [{"geq": ["ipTotalLength", 256]}]}, {"select": [{"less": ["ipTotalLength", 512]}]}]}]}, {"count": [{"and": [{"select": [{"geq": ["ipTotalLength", 512]}]}, {"select": [{"less": ["ipTotalLength", 1024]}]}]}]}, {"count": [{"and": [{"select": [{"geq": ["ipTotalLength", 1024]}]}, {"select": [{"less": ["ipTotalLength", 1500]}]}]}]}]}, {"add": [{"entropy": [{"count": [{"select": [{"less": ["ipTotalLength", 128]}]}]}]}, {"entropy": [{"count": [{"and": [{"select": [{"geq": ["ipTotalLength", 128]}]}, {"select": [{"less": ["ipTotalLength", 256]}]}]}]}]}, {"entropy": [{"count": [{"and": [{"select": [{"geq": ["ipTotalLength", 256]}]}, {"select": [{"less": ["ipTotalLength", 512]}]}]}]}]}, {"entropy": [{"count": [{"and": [{"select": [{"geq": ["ipTotalLength", 512]}]}, {"select": [{"less": ["ipTotalLength", 1024]}]}]}]}]}, {"entropy": [{"count": [{"and": [{"select": [{"geq": ["ipTotalLength", 1024]}]}, {"select": [{"less": ["ipTotalLength", 1500]}]}]}]}]}]}]},
  {"get": [14, "tcpControlBits"]}
"features": ["_KDD5", "_KDD23", "_KDD3", "_KDD6", "_KDD35", "_KDD1"]